For most of my career, building software was expensive.

You needed engineers. You needed time. You needed a roadmap. Even relatively simple internal tools could take months to scope, design, build, test, and deploy.

AI has completely changed that equation.

Today, I can look at a piece of software my company uses and think: I could probably build a version …

Read more

Beyond Anthropic, OpenAI did something interesting. Their flagship model split into three, a new wave of voice, and a quiet trend that matters more than any launch.

OpenAI’s GPT-5.6, now in three flavors

The headline drop is GPT-5.6, and OpenAI has stopped pretending one model fits everyone. It ships as Sol (frontier reasoning and long-horizon agent work), Terra (the everyday workhorse — roughly GPT-5.5-level quality at about half the cost), and Luna (fastest and cheapest). I like the honesty of this. Most of what I do at Bumo doesn’t need frontier reasoning; it needs “good enough, fast, and cheap enough that I stop counting.” Terra is the one I’ll actually live in. Sol is the one I’ll reach for when a task is genuinely hard and I’m willing to wait. The risk is decision fatigue: three names is two more than most people want to think about. The former SAT tutor in me makes up funny things to remember information: Sol (sun burns a lot of energy), Terra (Earth where you live every day), and Luna (the moon - bright but probably doesn’t burn a lot of energy).

BUT if the pricing gap is real, the tiering earns its keep.

Voice that doesn’t wait its turn

OpenAI also launched GPT-Live-1 and a mini version: full-duplex voice models, meaning they listen and talk at the same time instead of the walkie-talkie “okay, your turn” cadence we’ve all learned to tolerate. This is the release I’m most curious about. Interrupting an assistant mid-sentence and having it actually adjust is the difference between a demo and a thing you’d use while cooking dinner. I’ll believe the “feels natural” claim when I’ve talked over it a few times, but the direction is right.

Grok 4.5 and Gemini 3.5 Flash

Grok 4.5 landed a day earlier from Musk’s xAI. I’ll reserve judgment until the independent evals settle…Grok’s launch numbers and its lived experience haven’t always matched. More interesting to me was Google quietly making Gemini 3.5 Flash the default in AI Mode for everyone, globally. “Frontier-ish performance as the free default for billions of people” is a bigger deal than a new top-end model. Distribution beats benchmarks, and nobody has Google’s distribution.

The one that’s secretly about me

Base44 (the Wix subsidiary) shipped Base 1, its own in-house model trained on tens of millions of real app-building sessions. It’s the first “vibe-coding” platform to train its own model instead of renting someone else’s. That’s the move I’d watch if I were building any AI product with real usage: your users’ interactions are the moat, and eventually you stop paying rent and train on them. For those of us running lean, it’s also a reminder that the data you’re sitting on may be worth more than the features you’re shipping. As a serial Lovable user (since May 2025 - woot woot!), I am intrigued.

The actual story: price

Underneath all of it, CNBC reported that US companies are increasingly reaching for Chinese models like DeepSeek and Z.ai as OpenAI and Anthropic costs climb. Notice that both OpenAI (Terra, Luna) and Anthropic (Sonnet 5 leaning into cheaper, longer coding runs) spent this week emphasizing cost, not just capability. That’s not a coincidence. The frontier race got the headlines for two years; the affordability race is what will decide who actually gets embedded in the tools people use every day. As someone who pays these bills, I’m rooting for the price war.

What I’m taking from it

Capability is table stakes now. The interesting battleground moved to cost per useful task and how invisibly the thing slots into what you already do. Terra and Gemini Flash aren’t the flashiest things shipped this week; they might be the most consequential.

More next Thursday. PEACEEEE!!! ✌🏼

Sources

• Top AI News for July 2026 — AIapps

• July 2026 AI Releases — ThursdAI

• LLM News Today (July 2026) — llm-stats

• OpenAI’s GPT-5.6 models to be publicly released — Nextgov/FCW

• Google Search’s I/O 2026 updates — Google blog

• Chinese AI models gaining ground as US costs surge — CNBC

Sarah Willersdorf and I sat down and recorded an episode of Goop’s podcast, and instead of doing the usual founder/builder thing, we more or less took the mic and ran with it.

The topic: AI. Not the abstract, “in five years everything changes” version. The version that’s already here. The version I use to run a company, manage a household, and build things that would have taken a team of ten and a budget I didn’t have eighteen months ago.

If you’ve been feeling some mix of curiosity and low-grade dread about all of this, then this episode was made for you.

What we actually talked about

Sarah and I come at this from two different angles that happen to meet in the middle.

I’m building Bumo, and our AI family assistant Bennie (check it out here)is the first of its kind — it helps parents instantly book care and handle the logistics of family life. So I live in the “what can you actually ship right now”

question every single day.

Sarah spent years as Head of Luxury, Fashion & Beauty at BCG and as a CMO, and now advises boards and CEOs on where AI actually drives value versus where it’s just noise. So she lives in the “what should you bet on” question.

Put those together and you get the conversation I wish someone had handed me two years ago.

We got into:

• How we actually use AI day to day — from the unglamorous household admin to building entire companies. The real workflows, not the keynote version.

• Why women have a real competitive advantage right now. Communication, judgment, strategy, knowing what’s worth doing — these are the skills this moment rewards, and they’re skills a lot of us have been quietly compounding our whole careers.

• What builders are creating now that execution has gotten cheap. When the cost of making something drops to near zero, the bottleneck stops being “can I build it” and becomes “do I know what’s worth building.”

• What AI can’t replace. Human judgment. Conviction. The decision about what deserves to exist in the first place. The machine is extraordinary at execution. It has no idea what matters.

The line I keep coming back to

Execution is being democratized.

That sounds like a tidy panel soundbite, but sit with it for a second, because it changes everything.

For most of history, the people with the resources to execute… the capital, the team, the technical chop… held the power. The idea was cheap - the building was expensive. That gap is collapsing. SUPER FAST!

Which means the advantage shifts to the people who have taste. Conviction. The ability to look at a thousand possible things and say, “that one.” That’s not a technical skill. That’s a human one. And it’s the part that doesn’t get automated.

That’s also why I don’t think this moment belongs to the most technical person in the room. It belongs to the person who knows what to point the technology at.

Why I think the window matters

The thing I most wanted to get across in this episode: this moment is real, and it’s closing.

Not closing in a doom way. Closing in a “the people who move now will have a structural head start that’s hard to catch later” way. The tools are good enough, the cost is low enough, and the playing field hasn’t fully sorted itself yet. That’s the rare and brief overlap. It won’t stay this open.

So if you’ve been waiting for permission, or for things to settle down, or for someone to hand you a tidy roadmap — consider this the nudge.

If you want to actually start

The number one thing people tell me after a conversation like this is: okay, but where do I begin?

That’s exactly why Moj Mahdara and I built AI Maxxing — a six-week mentorship program for founders, operators, and executives who don’t want a lecture, they want to build. The first cohort starts September 3.

You can find it at joinaimaxxing.com — and because this is the Goop crowd, use code GOOP at checkout for 10% off.

Go listen

The full episode is live now. Sarah and I don’t pull punches on what’s hype and what’s real, and we get specific about how we’re each building this into our lives — work and home.

🎧 Listen to the episode on Goop’s podcast.

The moment is here. Let’s build!

— Joan

Your AI security team doesn’t need to be big. It needs to exist.

Everyone wants a formula. How many security people do I need? The honest answer is: it depends less on your headcount and more on how much AI is actually running inside your product and your company.

A rough starting point: if you’re under 500 people and AI isn’t core to your product, one person with 30–40% of their time dedicated to this is enough — for now. Not a new hire. Probably someone already on your engineering or security team who gets the problem and can own it.

If AI is your product, you need more. Three to five people who can cover the basics: trying to break your own systems before bad actors do, assessing what could go wrong with your models, and responding when something does.

The bigger your regulatory exposure — finance, healthcare, anything government-adjacent — the earlier and heavier you need to invest. Compliance and audit burden alone will eat people’s time before you’ve written a single security policy.

But here’s the thing most people miss: good tooling multiplies your small team. There are platforms now that watch what your AI systems are doing, flag weird behavior, and keep logs. One sharp person with the right tools can cover a lot of ground.

Under 500 people? Don’t build a team. Do these five things.

Trying to hire a dedicated AI security team at this stage is a waste of time and money. Here’s what actually matters:

First, find out what AI your company is already using. This sounds obvious. It never is. Employees are using ChatGPT, Claude, Copilot, and a dozen other tools you didn’t approve. Your product might have AI components that have quietly sprawled. Make a list. You cannot protect what you haven’t counted.

Second, your biggest risk right now is probably people copy-pasting sensitive information into AI tools. Customer data. Internal strategy docs. Financial information. It’s happening at your company today. A clear written policy and basic monitoring of what’s leaving your organization through AI tools is not glamorous, but it’s the highest-leverage thing you can do.

Third, read your contracts with your AI vendors. Does Anthropic, OpenAI, or whoever you use train on your data? How long do they retain it? This is in the fine print. You should know it cold.

Fourth, hire an outside firm to try to break your AI systems once a year. Red teaming — which just means paying professionals to attack your own systems before someone else does — is relatively affordable and doesn’t require any full-time headcount to benefit from. More on specific firms below.

Fifth, don’t overbuild. Custom detection systems, running your own AI infrastructure, a dedicated internal red team — that’s all later. Much later.

If you’re using Claude Code to manage AI agents, read this carefully.

This is the one that keeps security people up at night right now.

The problem with AI agents — systems where one AI model is directing others to take actions — is that mistakes compound. One bad instruction, one hallucination, one injected command buried in an email or document your agent read, and you have a chain of real-world consequences happening faster than any human can stop them.

Here’s what responsible use actually looks like:

Lock down permissions. Every agent should have access only to what it strictly needs. Read access to most things. Write access only to test environments by default. Live production credentials should never be in an AI agent’s hands unless you have very deliberate controls around that.

Build hard gates for irreversible actions. Sending emails to customers, touching live financial data, deploying code, making transactions — any action that can’t be undone needs a human to click approve. And that gate needs to be a real technical constraint, not a prompt telling the AI to “ask first.” Telling an AI to be careful is not a security control.

Contain the damage radius. Run agents in isolated environments. If something goes sideways, the harm should be stuck in a sandbox, not loose in your production systems.

Log everything. Every action, every decision. If you can’t answer “why did it do that” after the fact, you have a problem.

Assume anything an agent reads could be hostile. There’s an attack called prompt injection where malicious instructions are hidden inside documents, emails, or web pages your agent processes — instructions designed to override what you told the agent to do. This is real. Filter and validate external inputs before they reach your agent logic.

Set hard limits on what agents can spend and do per session. Runaway loops — where an agent just keeps calling itself or executing actions — are a genuine failure mode. A hard ceiling on cost or action count stops it cold.

The teams getting hurt right now are the ones who gave an AI agent full system access and figured the prompts would keep it in line. They don’t.

Start with read-only, low-stakes tasks. Expand access as you build real confidence. Treat it like onboarding a new employee — you wouldn’t hand someone the keys to production on day one.

So who should you actually call?

If you’re ready to hire an outside firm, here’s how to think about it based on where you are.

If you want software to do most of the heavy lifting — faster and cheaper:

Protect AI / Recon runs automated tests across hundreds of known attack types and can turn results around in hours rather than weeks. Good if you want broad coverage fast without a large engagement fee. Mindgard takes a similar approach but leans more toward continuous monitoring — it maps your AI attack surface and keeps watching it over time, useful if you want ongoing protection rather than a one-time snapshot. Gray Swan is more research-forward, backed by a large network of AI security researchers. A better fit if your AI systems are genuinely sophisticated and you want cutting-edge adversarial testing.

If you want humans running the engagement:

Schellman is a well-known name in compliance-heavy environments — they tailor their testing to your specific architecture and their engagements typically run around two weeks, starting at roughly $16,000. Worth it if you’re in a regulated industry or need something you can show an auditor. Shaip takes a notably human-led approach, bringing in domain experts and linguists alongside security analysts — good for catching the nuanced, contextual failures that automated scanning tends to miss. CrowdStrike is the bigger-brand option, better suited to mid-market and larger companies that want AI-specific testing added onto broader security work they may already be doing.

One important caveat before you call anyone: write down your two or three biggest fears first. Data leaking through an AI tool. An agent going rogue. A customer-facing chatbot saying something that creates legal exposure. Find a firm that specifically tests for those things. Don’t let a vendor define your risk profile for you — that’s how you end up paying for tests that don’t match your actual situation.

The bottom line

AI security isn’t a problem you can ignore until you’re bigger. The attack surfaces are forming right now, at every company size. But it’s also not something you need to spend a fortune on to get meaningfully safer.

Know what AI you’re using. Control what leaves your organization. Lock down your agents. Log everything. And get outside eyes on your systems at least once a year.

A note on expertise: I’m not a security professional, and nothing in this piece should be taken as formal security advice. These are observations and learnings gathered from conversations with people who do this work. If your company is thinking seriously about AI security — and it should be — please hire at least an outside consultant to assess your specific situation. What’s right for one company won’t be right for another, and a real expert will find things you missed.

There’s a line I hear constantly now from candidates: “Oh, I love AI. I use it all the time.”

I used to get excited when I heard that. Now it makes me ask a follow-up question.

Because when I dig in, “I use AI all the time” usually means: writing emails with ChatGPT, brainstorming ideas, cleaning up a scope of work or SOP. Maybe summarizing a document. These are fine habits. But they are not a competitive advantage. They are table stakes — the equivalent of bragging that you use email instead of sending letters by post. Nobody gets hired because they know how to use Gmail.

The bar keeps moving. And if you’re still standing at the 2023 version of “AI-savvy,” you are already behind.

What We Did at Bumo — And What It Actually Revealed

Starting in the summer of 2025, we changed how we hired at Bumo — across every single role. Customer service rep. Salesperson. Marketing. Didn’t matter.

Every candidate had to vibe code something on Lovable.

The assignment was simple: pick a real problem you care about and build something that addresses it. No templates handed to you. No one walking you through it. Just you, a tool you’ve probably never opened before, and a problem worth solving.

The point was never to see if they could code. It was to see something much more important: how someone behaves when they’re dropped into the unknown.

Could they figure it out? How long did it take them to start? Did they get curious, or did they freeze?

Here’s an example email from a candidate that was hired:

Dear Bumo Team,

These are all the deliverables on my end for the Generalist Role take-home assignment sent to me the other day. I had a lot of fun building and presenting it to you!

Also— once again thanks for taking the time to talk to me about the company on Thursday. I’ve been thinking about how I could contribute, especially by bringing structure to the operations or product side. Given my early career experience, I’m actually really interested in taking a hands-on position that supports the organizational side of things (honestly, I am interested to see how I’d work and grow in a fast start-up environment and taking on different roles)

If you’re open to it later, I’d love to ask questions and chat about the workflows, documentation, and the company growth overall.

For the take home challenge:

1. This is the link for my landing page: https://XXXXX.lovable.app/

2. 3-5 elevator pitch to my product:

I built OpenMarketFinder, an interactive landing page that helps people discover local farmer’s markets through a searchable, mobile-friendly directory. Users can filter by location, day of the week, and farm or craft specialties Each market has its own page with vendor lineups, schedules, and amenities, while vendors have profiles showing availability, product types, and contact info. The layout is designed to feel welcoming, intuitive, and highly usable across both desktop and mobile.

3. Here’s a quick walkthrough of the prototype and features on Loom:

https://www.loom.com/share/XXX

Would love to hear your thoughts! Hope to hear from you this week.

Best,

Candidate Name

We were also exploring N8N automations and workflows across the team — building real operational infrastructure, not just playing with prompts.

The clearest signal of a bad hire wasn’t the quality of what they built. It was the ones who kept delaying. Weeks would go by. “I haven’t had time.” “It’s been a busy week.” “I’ll get to it soon.”

That response told me everything. Because the entire premise of the exercise is that this is something that should save you time, not cost it. If someone fundamentally doesn’t understand that — if they see AI as an additional burden rather than a multiplier — they will never be an active AI operator. They will always be a passive user at best.

The Definition of “AI-Forward” Has Expired Twice Already

Here’s the uncomfortable truth for anyone reading this who considers themselves an AI power user: the definition keeps changing, and it changes fast.

In Q3 of 2025, an “AI-heavy adopter” meant:

• Running document analyses in ChatGPT

• Building custom GPTs

• Lightly prototyping in Lovable, Cursor, Replit, or V0

That was impressive — for 2025. It is not impressive today.

Right now, someone who is genuinely well-versed in AI is using Claude Code. They’re obsessive about it — not in a performative way, but in the way a craftsperson is obsessed with their tools. They’re not just prompting. They’re connecting things. They’re building an ecosystem around themselves: automations, workflows, integrations that eliminate entire categories of manual work. They are operating at a weight class that their job title and headcount don’t suggest.

They are, in every practical sense, punching way above their weight.

That’s the benchmark. And it keeps moving.

The Mistake of Thinking You Can Just Upskill Everyone

After the initial hiring push, we paused. We focused on upskilling the team we had rather than bringing new people in. And I want to be honest about something: I think that story is easy to misread.

It might sound like a case study in how you can transform any employee into an AI operator with enough training and encouragement. And to some extent, that’s true.

But there is a specific DNA required to become a genuine AI operator — and not everyone has it by default. You can coach someone’s skills. You cannot easily coach their relationship with uncertainty, their bias toward action, or their instinct to reach for a new tool when faced with a problem.

Some people, when dropped into Lovable with no instructions, immediately start clicking around. They break things. They figure out why. They ship something imperfect and feel good about it.

Other people — talented, smart, capable people — wait. They want to understand the full picture before they begin. They want to do it right. That instinct is valuable in many contexts. In AI adoption, it is a bottleneck.

The ones who thrive are the ones who already had that explorer’s orientation — the same DNA that makes someone learn to skateboard by falling a lot rather than watching YouTube videos first.

It’s a Two-Way Street — But the Finish Line Is in the Middle

None of this happens without leadership. A company will not become AI-forward if leadership just says “go figure it out” and retreats. You have to model it. You have to be genuinely obsessed with it yourself. You have to make it a cultural expectation, not just a nice-to-have.

But here’s the other side: a mandate without willing partners goes nowhere fast. The leader can run toward that finish line all they want — if the employees won’t move, you’re not meeting in the middle. You’re just standing at opposite ends of a very awkward road.

Think of it like a psychographic match. The leader who makes this work has a specific profile: curious to the point of obsession, comfortable with ambiguity, and more excited by what AI unlocks tomorrow than threatened by what it replaces today. The employee who thrives in that environment has a nearly identical profile — not the same skillset necessarily, but the same orientation toward the new.

When those two types find each other, the compounding effect is extraordinary. When they don’t, even the best tools and training programs end up collecting dust.

What to Look For — Practically Speaking

If you’re building a hiring process for this era, here are the signals I’d look for:

Green flags:

• They built something before you even followed up

• They describe a workflow they created, not just a tool they use

• They talk about what they connected, automated, or eliminated — not just what they prompted

• They’re slightly embarrassed by how they used AI six months ago because they’ve moved so much further

• They ask you what tools you’re using

Red flags:

• “I use AI for everything” with no specifics

• Delays in the take-home project framed as a time problem

• Describing passive use (summarizing, drafting, ideating) as if it’s a differentiator

• No curiosity about what you’re building internally with AI

• Treating the exercise as a test to pass rather than a problem to solve

The Bottom Line

Speed is commoditized now. Any reasonably competent person with the right tools can execute faster than ever before.

What is not commoditized is the quality of the idea, the instinct to reach for the right tool, and the ability to move from “I have a problem” to “I built a thing that addresses it” before most people have opened a new browser tab.

That instinct is what you’re hiring for. The tools are almost secondary — they’ll change anyway. What won’t change is whether someone has the explorer’s orientation that makes them genuinely dangerous in the best possible way.

Hire for that. Test for it directly. And don’t confuse enthusiasm for capability — the two are very different things.

Bumo started testing this hiring approach in summer 2025. The takeaways in this article come directly from that experience.

Not long ago, brand visibility online had a simple formula: rank on Google. Get on page one. Drive clicks. The game was well understood, the playbook was thick, and the agencies were expensive but predictable. Then large language models arrived — and quietly, almost surgically, they started answering questions instead of listing links. The old playbook dissolved almost overnight.

Today, when someone asks ChatGPT, Perplexity, or Google’s AI Overviews which project management tool is best for a remote team, they don’t get ten blue links. They get an answer. And that answer is assembled from somewhere. The question every brand now faces is brutally simple: are you the source, or are you invisible?

This is the domain of Generative Engine Optimization (GEO) — also called Answer Engine Optimization (AEO) — a discipline that barely had a name two years ago and is now arguably the most consequential shift in digital marketing since the advent of SEO itself.

What GEO Actually Means

Traditional SEO optimizes content so search engine crawlers rank it highly in results pages. GEO optimizes content so that AI language models cite, reference, or synthesize it when generating answers. The distinction sounds subtle; the implications are enormous.

With SEO, a brand earns a slot on a list. With GEO, a brand becomes part of the answer. There are no “positions two through ten” in an AI-generated response — there’s the answer, and then there’s everything the model didn’t mention. The winner-take-all dynamics are more extreme than anything traditional search ever produced.

The factors that drive GEO are also fundamentally different. Search engine algorithms prioritize backlinks, keyword density, and technical structure. LLMs prioritize trust signals, conversational authenticity, community consensus, and what researchers have started calling “citation gravity” — the property of content that makes AI models reach for it when assembling an answer. And for most of 2025, one platform had citation gravity in abundance.

The Reddit Era: Why It Dominated

Reddit’s rise as the backbone of AI-generated answers wasn’t accidental. It was the product of structural decisions — by Reddit, by AI companies, and by Google — that converged into one of the most unusual chapters in platform history.

In early 2024, Reuters reported that Reddit signed a $60 million per year contract with Google, granting the search giant access to Reddit content for training its AI models. OpenAI separately paid for access to Reddit’s Data API. Almost simultaneously, Google’s algorithm updates sent Reddit’s search traffic surging from roughly 500 million to 3.4 billion monthly visits in a single year. The effect on AI citations was seismic.

By June 2025, a Semrush analysis of over 150,000 AI citations found that Reddit was the most cited domain across LLM responses — accounting for 40.1% of all citations, ahead of Wikipedia, YouTube, and every major publisher on the web. For brands trying to appear in AI answers, the message seemed clear: get on Reddit.

The reasons for Reddit’s dominance were structural, not accidental. Reddit conversations mirror how people actually ask questions — messy, specific, emotionally inflected. They’re rich in first-person experience, comparative judgment, and the kind of nuanced opinion that polished corporate content can never authentically replicate. When someone asks an AI “is Notion actually worth switching to?”, the answer that feels most useful comes from a real person who switched — and those people live on Reddit.

The data licensing deals added another layer: because Reddit content was used to train LLMs directly, the models didn’t just cite Reddit — they were, in a sense, shaped by it.

The YouTube Reversal

Then, beginning in the second half of 2025, something shifted.

On January 26, 2026, Adweek published data from four independent research firms — Bluefish, Emberos, Goodie AI, and Profound — confirming a major reversal in AI citation patterns. YouTube had overtaken Reddit as the most-cited social platform in AI-generated responses. YouTube now appears in 16% of LLM answers; Reddit has fallen to 10%. Reddit’s stock dropped 9.3% following the report.

The mechanism behind the shift is instructive. For years, YouTube’s enormous library of video content was largely inaccessible to AI models — you can’t easily parse a video the way you parse text. But YouTube’s infrastructure of auto-generated transcripts, chapter markers, timestamped descriptions, and creator-written explainers created a parallel text layer that LLMs could read. Once AI systems got better at processing that layer, YouTube’s vast archive of authoritative, long-form expertise became fair game.

The structural reason YouTube now wins where Reddit once dominated comes down to one word: authority. Reddit’s pseudonymous, upvote-driven model makes it hard for AI systems to evaluate who is actually speaking. A three-year-old comment from a throwaway account carries the same formatting weight as expert advice. YouTube, by contrast, connects content to a named creator with a track record, subscriber count, and history — signals that AI models increasingly use to weight credibility.

The Nuance the Headlines Miss

Here is where the story becomes genuinely complex — and where brands that act on headlines alone will make expensive mistakes.

The “YouTube overtakes Reddit” narrative captures a real shift in percentage share of social citations. But first-party tracking data from Superlines, monitoring 62 brands over the 30 days ending February 2026, tells a different story in absolute terms: Reddit still generated 39,551 citations compared to YouTube’s 15,735 in the same period.

More important still is how Reddit is being cited. Research from Tinuiti and Conductor shows that while Reddit’s overall citation frequency dropped roughly 50% between October 2025 and January 2026, Reddit’s citation share in commercial categories — technology, electronics, product comparisons — actually grew 73%. And sole-source citations (responses where Reddit is the only reference) rose 31%.

Reddit is becoming more selective, not less influential. LLMs are moving from quantity-based sourcing to intent-based citation. For the specific prompts that matter most to purchasing decisions — “should I switch from X to Y?”, “what’s the best tool for Z?” — Reddit still wins by default because no other platform generates that kind of authentic peer testimony at scale.

The platform-by-platform variation adds another layer of complexity. January 2026 data shows ChatGPT cites Reddit in more than 5% of responses; Perplexity draws 31% of its citations from social media with Reddit dominant; Gemini cites Reddit in just 0.1% of responses. A brand that monitors only its Gemini presence may be entirely unaware that ChatGPT is assembling its product evaluation from Reddit threads it has no visibility into.

What Brands Must Do Now

The old GEO playbook — “get mentioned on Reddit, win AI search” — was always a simplification. The new landscape demands something more sophisticated: a multi-platform, multi-format strategy built around the fundamental question AI models are asking about every piece of content they encounter. Is this trustworthy? Is this specific? Is this the kind of thing a real person in this situation would actually say?

1. Build a YouTube Knowledge Layer Create structured video content with manually edited transcripts, descriptive chapter markers, and keyword-rich descriptions. AI models parse the text layer around videos — not just the video itself. Tutorials, comparisons, and product deep-dives with clear timestamps create what researchers call “copyable atoms”: discrete, citeable units of information that LLMs can lift cleanly into an answer.

2. Treat Reddit as Reputation Infrastructure Don’t abandon Reddit — concentrate your presence where it matters. Engage in the subreddits where purchasing decisions happen in your category. Use thought leaders and real employees, not brand accounts. The goal is not volume of mentions but depth of positive, specific, upvoted experience reports that LLMs will surface when someone evaluates your product.

3. Audit by Platform, Not by Channel Because ChatGPT, Perplexity, Gemini, and Google AI Overviews each source differently, your brand’s AI visibility may be radically inconsistent across them. Run regular audits — testing the same prompts across each major LLM — and treat each platform as a distinct visibility channel requiring its own monitoring strategy.

4. Invest in LinkedIn Authority LinkedIn’s emergence as a surprise citation contender — nearly matching YouTube’s citation volume in early 2026 — signals that professional context and named expertise are increasingly valued by LLMs. Long-form posts from genuine practitioners, published consistently and linked to a clear domain of expertise, are building serious citation gravity.

5. Structure Everything for AI Parsing Regardless of platform, content that wins AI citations shares structural traits: clear headings, numbered steps, specific data points with dates and sample sizes, and natural question-and-answer framing. Avoid blocking AI crawlers (GPTBot, CCBot, ClaudeBot) in your robots.txt. Ensure linked pages load without heavy JavaScript.

6. Make Citations a KPI Stop measuring only traffic and rankings. AI search traffic converts at 14.2% versus Google’s 2.8%, according to available benchmarks. Track how often your brand appears in AI-generated answers, on which platforms, in what sentiment, and compared to competitors. Tools like Semrush, Superlines, and Profound now offer AI citation tracking at scale.

The Harder Truth About GEO

There is a version of the GEO story that flatters marketers: optimize your content, show up in AI answers, grow revenue. It’s clean, actionable, and sellable. But the harder truth is that generative engine optimization is, at its core, a reputation problem — not a content problem.

LLMs surface Reddit threads with equal authority whether those threads are praising your product or savaging it. Research shows that citation rates for positive and negative brand sentiment on Reddit are nearly identical. LLMs are not filtering for constructive feedback. They are surfacing community consensus — and if that consensus is negative, no amount of technical optimization will change what an AI tells your prospects.

The brands that will win the GEO era are not necessarily those with the most optimized content structures. They’re the ones whose genuine reputation — the accumulated weight of real customer experience, published at scale, across every platform an AI might consult — is strong enough to pull AI answers in their direction without needing to game anything.

That’s a higher bar than traditional SEO ever demanded. It’s also a more durable one. Algorithms change. Search parameters shift. The Reddit-YouTube inversion happened in under six months. But a genuinely trusted brand — with real practitioners publishing real expertise on YouTube, real customers sharing real experiences on Reddit, and real thought leaders building real authority on LinkedIn — is not one algorithm update away from invisibility.

The Bottom Line

GEO is not a replacement for SEO — it’s an additional layer of the same fundamental question: when someone needs what you offer, do they find you? The stakes are higher now because AI-generated answers increasingly eliminate the click entirely. Visibility is binary. You’re in the answer or you’re not. The brands building for that binary future — across Reddit, YouTube, LinkedIn, and every other platform an LLM might trust — are the ones who will own the next decade of search.

I didn’t want to lecture. I wanted to ignite. So I built a framework and experience that would do just that.

💡 Why "Pace with Purpose"?

This became our North Star. The idea is simple: our goal isn't to just move faster or work longer. It’s to work intentionally. That means making time count by building core productivity habits and multiplying output with smarter tools like AI and automation.

I told my team: “You should be angry that manual work is stealing your time—the time you could be using to rest, reflect, or solve higher-order problems.”

This wasn’t just a productivity pep talk. It was about valuing ourselves and our time deeply enough to demand better workflows.

🔧 The Framework: Foundations + Multipliers

1. The Foundation Layer: Productivity Drivers

Before even touching AI, I wanted the team to master the basics. These are the pillars of purposeful work:

• Time – Be present, accountable, and thoughtful about how time is allocated. And invest the time!

• Clarity – Know exactly what needs to get done, why, and by when.

• Focus – Shut out the noise. Deliver on what matters.

• Organization – Create clean, streamlined systems so you can move without friction.

When these habits are strong, adding efficiency tools like AI doesn’t just boost productivity—it explodes it.

2. The Accelerator Layer: Efficiency Multipliers

Once the foundation is solid, here’s where we accelerate:

• Automation – From simple task flows to more advanced integrations, anything repetitive should be delegated to systems.

• AI – Tools like ChatGPT and Notebook LM become trusted collaborators for planning, writing, brainstorming, summarizing, and decision-making.

But only if we really learn how to use them—not just superficially.

⚡ Electrifying the Training Experience

To bring this to life, I didn’t just hand over a slide deck. I built a mini AI adoption ecosystem:

• 🎧 Podcasts via Notebook LM – I recorded short, energizing sessions explaining the mindset shift and use cases for AI. These weren’t just training materials—they were designed to inspire. Listen to the “Podcast” that was created in 3 minutes:

• 🗂️ Slack Canvas & Slack Lists – Training materials, progress tracking, and AI toolkits were embedded right in our communication flow. That reduced friction and created shared accountability.

  

This wasn’t a one-off. It became an evolving system of support that employees could tap into anytime, anywhere.

🧠 Upskilling as a Team Sport

Using AI isn’t about one shiny tool or one power user. It’s a muscle—and everyone needs to build it.

What made this harder is that my team is smaller than most people would guess. That’s why it was even more critical for every person to level up.

I wasn’t just training for knowledge—I was developing autonomy, speed, and critical thinking. And when even one team member becomes AI-fluent, the whole team benefits from their increased leverage.

✅ What It Means to Be Productive at Bumo

At Bumo, we define productivity like this:

1. Start with strong fundamentals (clarity, focus, time, organization).

2. Multiply your output with automation and AI.

3. Ask constantly: “How can I work smarter—not just longer?”

This isn’t a passing phase. It’s a cultural shift. And once people feel the difference—when a task that used to take 3 hours now takes 30 minutes—it becomes addictive in the best way.

So if you want your team to adopt AI, don’t start with the tool. Start with the why. Then show them what’s possible. Visually. Practically. Electrifyingly.

Let them feel the power of valuing their time.

Leave a comment

Well, the future just gave that idea a major upgrade—and your kids are the headliners.

This year, for Father’s Day, we made something unforgettable using Suno.com—a wild new AI tool that turns anything into a full-blown, original song. Lyrics, melody, production—all handled by AI. You just give it the words, and Suno gives you a studio-quality track. No music skills needed. No garage band required.

🎤 The Kids Took the Mic

We asked each of the kids one simple question:

"What do you love about Dad?"

Here’s what they said:

Elliott, age 7: "He picks me up from school. He’s funny. He tells me stories about Barry (our dog that passed)."
Reagan, age 8: "He’s the best. He is funny and cuddles with me.”

We dropped those lines into Suno. Thirty seconds later: boom. A jam was born. I asked it to write a song about how much he’s always there for the kids. We lost our home and the kids lost their school in the Eaton fire, and he’s really been our rock.

“Strong for the Kids” by Joan AI Records:

🪄 From Sweet Words to Sweet Sounds

Here’s the best part: it felt just like those mix CD days, but better. Instant. Personal. Surprise-filled. And if I could burn it onto a CD, I absolutely would. Sharpie label and all.

💡 How You Can Do It Too

1. Go to Suno.com

2. Sign in with your Google or Discord account

3. Click Create, enter your lyrics or thoughts

4. Choose a style (Pop, R&B, Acoustic, etc.)

5. Hit Generate

6. Cry/laugh/dance

You can download the song, share it, even post it to Spotify if you want. It’s like gifting a song written by your kids—but with Beyoncé-level production.

This Father’s Day, skip the tie. Skip the mug.
Give Dad something you literally couldn’t have made a year ago.

A song. From the kids. About him.
Because he deserves a mixtape from the future.

My headshot is… old. Like, “still wearing glasses I don’t even own anymore” old. It just doesn’t look like me. So instead of booking a photographer or even asking a friend with Portrait Mode, I did the very normal thing: I used AI to make a version of me… without me.

We’ve seen the yearbook AI craze. We’ve seen the ultra-glam avatars, the futuristic edits, the slightly uncanny versions of ourselves that are equal parts impressive and mildly terrifying. And naturally, I had to try it.

I tested three tools. Each one promised to help me look more like me—just without the whole “get camera ready” part. Spoiler: I had to use two apps just to get one image that felt right. Here’s the breakdown:

Tool #1: Photoleap - The “Wait, Is That a Floating Head?” One

Photoleap (photoleapapp.com)– A creative and powerful photo editing app by Lightricks, offering AI tools for quick edits to pro designs.

Key Features:

• AI-powered tools for background removal, object removal, and image enhancement.

• Text-to-image generation capabilities.

• A wide array of filters, effects, and animation tools.

• Layer-based editing for complex compositions.

Pricing:

• Free Version: Access to basic editing tools.

• Pro Subscription:

  • Monthly: Approximately $6.99/month.

  • Annual: Approximately $59.99/year.

• Free Trial: 7-day free trial available for new users.

  

Let’s just say this one didn’t make the cut. The head-to-body ratio was… off. It looked like a botched Photoshop job from 2008 where someone tried to paste a celebrity’s face onto a stock image. It also gave me lots of wardrobe choices—yay!—but it didn’t matter because every version just looked fake. Not AI-futuristic-fake. Just fake-fake. And my face was captured in the same angle in every photo.

Tool #2: Glam.- The “Supermodel If You’ve Never Met Me” One

GlamAI (glam.ai) – An AI-powered video and photo editing tool that enhances your content with a variety of styles, from AI art filters to anime and popular social media trends.

Key Features:

• AI styles including Anime, Rock Star, and Neon.

• Digital wardrobe feature to change outfits in photos and videos.

• Regular updates with new styles and effects.

• Video editing capabilities with AI enhancements.

Pricing:

• Free Version: Limited access to basic features.

• Pro Subscription:

  • Weekly: $7.99/week.

  • Monthly: $22.49/month.

  • Annual: $64.99/year.

• In-App Purchases: Additional styles and features may require separate purchases.

Glam’s vibe? Black-and-white, artsy, and gorgeous—but not practical. These weren’t headshots; they were giving album cover meets lifestyle magazine. Also, while the images looked fairly realistic, they clearly weren't designed for a 5’3” size-6 human like me. If you’ve never met me, maybe you’d believe the tall, ethereal woman it generated was me. But anyone who knows me would call Photoshop on sight.

Tool #3: Momo - The “Almost Got It” One

Momo AI (momoai.com)– An AI photo editor that effortlessly customizes photos to align with your professional goals, helping you create impeccable AI photos on the fly.

Key Features:

• AI training using 8-12 personal photos to create customized images.

• Various photo styles including formal, casual, and themed.

• Tools for noise reduction, face enhancement, and image upscaling.

• High-resolution image outputs suitable for professional use.

Pricing:

• Free Version: Basic features with limited customization.

• Pro Subscription:

  • Weekly: $6.99/week.

  • Monthly: $29.99/month.

  • Annual: $39.99/year.

• In-App Purchases: Coin-based system for purchasing additional styles and features.

Momo actually came closest. The lighting was great, the expressions looked natural, and it felt like me—kind of. But it defaulted to making me look really thin. Like... she-does-Pilates-and-drinks-green-juice thin. So I had to do something wild: I used Facetune—not to touch up, but to touch down.

Yes, I used Facetune to reverse-enhance myself. I gave myself less pouty lips, a wider jawline, a shorter neck, and made myself look more like my actual size. And then I brought it into Canva, used some expansion tools to widen the background (because AI apparently thinks we all live in tight corners), and finally got something that said:

"Hi, I’m Joan. I’m human. I’m real. And yes, I did make this headshot without ever leaving my couch."

My causal images made with AI.

Here are more business-y ones!

It’s like the assistant you always dreamed of—effortless, available 24/7, and shockingly intuitive when it comes to managing your schedule, juggling meetings, and keeping your calendar clean and current.

Below are some of the very real, very satisfying ways it has made my life smoother:

📅 Coordinate Meetings Like a Pro

Need to book a time with one person? Or five? Howie’s got it. Just loop him in and watch the magic unfold.

• "CCing Howie to find us 45 mins to meet in May"
  → No back-and-forth needed. Just results in your inbox.

• "Howie, please reach out to these 4 people and schedule 30-minute calls next Mon/Tue"
  → Group scheduling? Delegated and done.

• "Howie, can you grab a time on Erica's Calendly?"
  → He’ll go to her link, pick a time that works for you, and confirm it.

🗓️ Add Events to Your Calendar—Without Lifting a Finger

Whether it’s a flight, a board meeting, or school commitments, you can just email Howie and move on.

• "Here is my upcoming PTA meeting schedule. Please add all of them to my calendar"
  → One email and you're organized for the semester.

• "Find me 2 hours to work on the board deck before Wednesday EOD"
  → He’ll look at your availability and hold time accordingly.

• "That sounds great. Howie will send an invite."
  → Just drop it into the thread. He takes it from there.

🔁 Interact with Existing Meetings Without Breaking a Sweat

Need to move something? Add a person? Add a buffer to make your life easier? Howie can handle it all.

• "Howie, please reschedule all of my meetings tomorrow."
  → Weather emergency? Family crisis? He’s got your back.

• "Can you add Dave to the sales call on Monday?"
  → No calendar access needed—just a quick email.

• "Put 'call Uber' on my calendar 2 hours before my flight out of SFO this weekend"
  → Howie understands time zones, weekends, and travel.

💡 Why Howie Works So Well

What’s brilliant about Howie is that he doesn’t try to do everything. He focuses just on scheduling—and he nails it. A few things that make him different:

• 100% Email-Based: No new apps. Just email him like a real assistant.

• Knows Your Calendar: Once connected, he sees your availability and preferences across all calendars.

• Natural Language: Talk to Howie like you would an EA. No special commands.

🧠 Bonus Use Cases I Love

• "Howie, decline all Friday meetings after 1pm moving forward."

• "Can you send weekly invites for team syncs through August?"

• "Reschedule my 2pm if the board prep isn't done by noon."

• "Follow up with Emily if I don’t hear from her by Thursday."

Yes, Howie can remember, track, and execute those too. It’s almost scary. In a good way.

🧪 Still in Beta, Still Blowing My Mind

Howie’s in beta—but the pace of improvement is nuts. Every few weeks he gets smarter: better memory, more context awareness, and a sharper eye for what you need before you ask for it.

Join Howie's Waitlist

If you’re drowning in calendar coordination or constantly rescheduling things on the fly, I can’t recommend Howie enough.

It’s like hiring an assistant—but without the onboarding.

I don’t like shopping. I should work out but don’t. Also, I have not one athletic bone in my body. So, I spend my time exploring AI tools and have too much fun doing so. Like, stay-up-way-too-late kind of fun. Every single day, there are dozens of new AI tools launching — but let’s be real: most of them kind of suck.

Every time I search for new tools, it feels like scrolling through the Cheesecake Factory menu. An overwhelming directory of options, but no clue where to start, what’s actually good, or what’s worth my time.

Inside my company Bumo (www.bumo.com), we even had a Slack channel dedicated to AI tools in order to engage our team and get everyone to use AI. People were posting non-stop. But most of us weren’t really using these tools we listed. Or if we did use them, it was not in a way that actually changed how we worked. And honestly, it’s overwhelming and you’re trying to develop a muscle that feels unnatural. Even for the most powerful tools like ChatGPT or Claude — 99% of people are barely scratching the surface.

Here’s what I’ve realized: when you actually use AI — like really lean in — it can be transformative. I had ChatGPT walk me through running a Python script in a few minutes in place of creating custom coupon codes on Stripe manually (would have taken me an hour). But more than saving time, I felt empowered. Like, "I can build anything" empowered.

So that’s why I’m here.

On this Substack, I’ll share:

• The tools I actually use.

• Exactly how I use them.

• Whether they're worth your time (or not).

• The good, the bad, and the totally overhyped.

Some tools automate my scheduling — they literally book my meetings, follow up if someone ghosts me, and even schedule my son’s playdates. Others just sneakily make my life 10x easier behind the scenes.

It’s not just about AI-native tools either. So many of the apps we already use have quietly started layering in AI — some brilliantly, some embarrassingly bad. I’ll break those down too.

I invite you in. If you’re a founder, an operator, a creative, a parent, let’s empower ourselves together.

Subscribe now

“What a time to be alive” is my default response to someone showing me a new AI tool or platform. Just a few years ago, designing a simple graphic on Adobe Illustrator would take hours (and I have been using Adobe suite since I was 14, er, 26 years ago). Compared to the way I design now, Adobe was a timekill and a pain in the you-know-what.

Today, with AI-powered tools like ChatGPT and Canva AI (still can be a few notches better IMHO), what once required painstaking effort can now be accomplished in minutes.

AI is no longer just a futuristic concept; it’s like the thing you do and should do every day. I use AI more times in a day than I drink water (terrible but true). Platforms like Canva’s WYSIWYG (What You See Is What You Get) design tooling allow users to quickly generate high-quality visuals, leveraging built-in brand elements to maintain consistency. Canva has the Brand Kit and Brand Guidelines tool that has been so incredibly helpful to my team. 1

At the same time, ChatGPT assists with generating copy wevn if yw type like thds it will be ableeee to figwer out what you are sying. You can brainstorm ideas, “How should I organize this deck” or “What copy should I use to promote XYZ campaign for ABC audience on 123 platform?” I can ask ChatGPT a questions as if I am running on no sleep with my contact lenses out (I am legally blind) with snow gloves on and it will still decode my partially coherent thoughts into articulate and thoughtful ideas. (If Homer Simpson had AI, he could have become president.)

The combination of these AI-powered tools has dramatically streamlined the creative process, making high-level design and content creation so accessible. Many of my team members who use Canva are not graphic designers by trade. The rapid upskilling because of AI and AI-powered tools is insanity.

Beyond creativity, AI is also transforming strategic thinking. At our company, we encourage our team to "punch above their weight" by leveraging AI as a tool for sharper decision-making. It acts as an instant knowledge repository, helping individuals answer their own questions, find the best resources, and serve as their own sounding board. You can great CUSTOM *cue the harp music* GPTs. YESSIR! We are feeding information into our custom GPT to support our team members so they can ask GPT. “Ask GPT. Don’t ask me” will be our new mantra for basic questions.

This self-sufficiency fosters a more proactive and empowered workforce, capable of thinking through innovative strategies with the assistance of machine learning models. TLDR; We are baddies because of AI.

The obsession with AI is understandable. It represents an era of truly and wildly unprecedented efficiency and accessibility, opening doors that were once closed to all but the most skilled professionals. Yet, as we integrate these tools into our lives, we must remain conscious of the need to use AI as an enhancer rather than just a total crutch. The future belongs to those who learn to wield AI wisely—using it not to replace human creativity, but to amplify it.

As AI continues to evolve, so must our approach to it. The key is not just to embrace AI but to master it—harnessing its power while maintaining our unique human edge. The real innovation lies not just in the technology itself, but in how we use it.

Mic drop.