The short version: to earn yield on stablecoins, you swap into a stablecoin that generates yield at the token level, then supply it and hold the receipt. Through Sky.money, that means trading USDC, USDT, ETH or SKY for USDS, supplying the USDS, and receiving sUSDS — a yield-generating stablecoin that accrues continuously and can be redeemed at any time, with no lockup.

Subscribe now

That’s the whole thing. The rest of this article is why it works that way, and what to check before you do it.

Why idle stablecoins are the most common unforced error in crypto

A stablecoin balance sitting in a wallet is doing exactly one job: holding a dollar. That’s a real job. It’s just not the only thing that dollar is capable of.

The reason most people leave them idle isn’t laziness — it’s that the alternatives look like work. Bridging, approving, monitoring a position, watching a rate move with borrowing demand, remembering to claim something. If putting your stablecoins to work means acquiring a small part-time job, most people correctly decide it isn’t worth it.

So the useful question isn’t where is the highest number. It’s which mechanism requires nothing from me after the first five minutes.

The three ways to earn yield on stablecoins (and how they differ)

Almost every option in the market is one of three shapes. They fail in different ways, which matters more than what they pay.

1. Lending markets. You supply a stablecoin to a pool, borrowers pay to take it out, you receive a share. The rate floats with utilisation — when borrowing demand is high, so is the rate; when it’s not, it isn’t. You are exposed to the solvency of the borrowers and the accuracy of the liquidation machinery.

2. Liquidity provision. You put a stablecoin pair into a pool and collect a share of swap fees. Stablecoin pairs limit impermanent loss, but your rate is a function of trading volume, which nobody controls and nobody can forecast.

3. Yield at the token level. The yield is embedded in the token itself. You hold it; it accrues. There’s no position to manage, no counterparty you selected, nothing to claim. This is the category sUSDS sits in, and it’s the one that survives being forgotten about.

None of these is universally correct. But if the thing you actually want is a dollar that doesn’t sit still, category three is the one designed for that.

What a yield-bearing stablecoin actually is

Here’s the mechanic that most explanations bury.

sUSDS is an ERC-4626 vault token. When you supply USDS and receive sUSDS, your sUSDS balance stays a fixed number. What changes is how much USDS each unit of sUSDS is worth. It accrues continuously — every block, in the background, with no action from you.

That single design choice removes most of the work:

• No claiming. There’s no rewards tab, no pending balance, no gas spent harvesting.

• No lockup. No unbonding period, no notice window, no withdrawal queue. Redeem whenever.

• No position management. There’s no health factor, no liquidation price, nothing to top up at 3am.

• No rebasing surprises. Your token count doesn’t change, which keeps accounting and tax records legible.

When you’re done, you redeem sUSDS for USDS plus everything it accrued. Fully liquid, both directions.

How to swap USDC or USDT for a yield-bearing stablecoin

The practical sequence, start to finish:

1. Go to app.sky.money and connect your wallet. Sky.money is a non-custodial interface to the decentralised Sky Protocol. It never holds your funds — every step is a transaction you sign yourself.

2. Open Trade. Select the token you’re holding — USDC, USDT, ETH or SKY — and select USDS as the token to receive.

3. Read the quote. This routes through permissionless, non-custodial pools, so you’ll see the expected output and price impact before you sign. Read it rather than assume it. USDS is engineered to absorb this kind of flow at scale through the Peg Stability Modules, so depth on major stablecoin pairs is deep — but the number on screen is the number that matters.

4. Approve, then confirm. You now hold USDS.

5. Supply your USDS. Open the savings widget, enter the amount, sign.

6. You hold sUSDS. It starts accruing immediately.

Two or three signatures depending on approvals. The only real variable is gas.

To exit, run it backwards: redeem sUSDS for USDS, and trade back out if you want to.

Where the yield on sUSDS comes from

This is the question worth being fussy about.

Yield that you can’t explain is yield you can’t assess.

The yield on sUSDS is the Sky Savings Rate — governance-set, and verifiable onchain. The chain of causation:

1. Sky Protocol makes USDS liquidity available to the Sky Agent Network.

2. Agents borrow that USDS and deploy it into yield strategies — collateralised loans, US Treasury bills, lending market liquidity.

3. Returns flow back to Sky Protocol.

4. Sky Protocol governance sets the Sky Savings Rate parameters based on aggregate performance.

5. sUSDS holders receive it automatically.

Two consequences follow, and they’re the two things worth internalising before you do anything:

The rate is governance-set, not utilisation-driven. It doesn’t spike and collapse with borrowing demand on a curve. It’s a parameter voted on by Sky Protocol governance. That makes it predictable in a way most onchain yield isn’t.

Predictable is not the same as fixed. Governance can move it. If you supply today and the rate changes next quarter, your sUSDS keeps accruing — at the new rate. Anyone describing a stablecoin yield as guaranteed is describing something other than reality.

AUTHOR’S NOTE

I’m deliberately not printing a percentage in this article. Rates move; articles don’t. Check the live figure on Sky.money before you decide anything.

What to check before you supply

The unglamorous list. Read it.

• Jurisdiction. Some Sky Protocol functionality accessed through the Sky.money interface — including the Sky Savings Rate and Sky Token Rewards — isn’t available in selected jurisdictions, including the U.S. The Trade feature isn’t available in EU member states. Check the Terms of Use for your situation before you connect, not after.

• Gas economics. On Ethereum mainnet, a small balance can spend more on three transactions than it accrues in months. USDS and sUSDS are live on Base, Solana and Arbitrum via SkyLink, where that arithmetic looks very different. Match the chain to the size of the balance.

• Smart contract risk is never zero. Sky Protocol has seven years of operation with zero exploits behind it. That’s a real track record. It is not a guarantee. Both things are true simultaneously.

• Self-custody means self-responsibility. Nothing in this flow is custodial. You keep control of your keys the entire time, which is the point — and the obligation.

• Records. In most jurisdictions every swap and every redemption is a taxable event. Keep the transaction hashes. Filing-season you will be grateful.

Frequently asked questions

Q. Can you earn yield on USDC directly?

A. Not natively at the token level. USDC doesn’t generate yield by being held. To put it to work onchain you either supply it to a lending market, provide liquidity, or trade it for a stablecoin that generates yield at the token level — which is the route described above.

Q. What is a yield-bearing stablecoin?

A. A stablecoin whose value accrues without any action from the holder. sUSDS is an ERC-4626 vault token: the balance stays constant while the redeemable value per unit rises.

Q. Do you have to lock up stablecoins to earn yield?

A. Not with a token-level mechanism. sUSDS has no lockup, no unbonding period and no withdrawal queue — it can be redeemed for USDS at any time.

Q. How often does sUSDS compound?

A.Continuously. Accrual is built into the token’s exchange rate rather than paid out on a schedule, so there’s nothing to claim and no compounding decision to make.

Q. Is the yield fixed?

A. No. The Sky Savings Rate is set by Sky Protocol governance and can change. It’s designed to be predictable rather than volatile — that’s not the same as fixed.

Q. What happens to accrued yield when you redeem?

A. It comes with you. Redeeming sUSDS returns USDS equal to your principal plus everything accrued to that point.

The one-line version

If you’re holding idle stablecoins and you’d rather they weren’t idle: trade into USDS, supply for sUSDS, and then genuinely forget about it. The design intent is that there’s nothing left to do.

Boring, on purpose. That’s the feature.

DISCLOSURE

A note on what this is: these are my personal views, written in my own capacity. Nothing here is investment, financial, legal or tax advice, and I am not a licensed adviser of any kind. I’ve described how the Sky.money interface and Sky Protocol work as I understand them, but mechanisms change, rates change, and availability varies by jurisdiction. Verify everything against the official documentation and live data before you act, and do your own research. Digital assets carry risk, including the risk of loss.

Then Stream Finance disclosed an external manager had lost roughly $93 million of assets. Deposits froze. xUSD fell toward thirty cents.

The interesting part is what came next. Only one of roughly 320 vaults on Morpho held xUSD directly, producing about $700,000 of bad debt. A rounding error. Yet queues formed across vaults that never touched it.

Nobody who supplied to those vaults had been told they were exposed to a fund manager they had never heard of. They were told a number.

That is the failure mode here. Not fraud. Not a hack. A number, presented without the mechanism that produced it.

Subscribe now

The rate is a summary statistic, and summary statistics lie by omission

APY is a compression algorithm. It squeezes five facts into one scalar: who sets the rate, what funds it, what breaks it, whether you can leave, and whether you can verify any of it.

Two products can print the same number while differing on all five.

Here is the decompression. Five questions, ten minutes per product. I use them on everything, including products I work on.

Question one: who sets the rate?

Not who pays it. Who decides it. Every stablecoin yield has a rate-setter, and which kind it is matters more than the number.

• A market sets it. Aave and most lending venues price by utilisation. Borrowing demand rises, your yield rises. Nobody voted. The curve did.

• A derivatives market sets it. Ethena’s sUSDe captures funding paid by long perpetual traders. Its seven-day average was 9.4% in late April 2026 against a ninety-day average of 11.8%, and an all-time low of 4.1% during the August 2024 funding inversion. That range is not a flaw. It is the product.

• A central bank sets it. Tokenised Treasury wrappers like USDY and BUIDL pass through a short-duration government coupon minus a fee. Your rate-setter is the FOMC.

• A governance process sets it. The Sky Savings Rate is set by SKY governance through an executive vote, calibrated against Sky Protocol revenue. It is not a utilisation curve, not algorithmic, and it does not track the FOMC mechanically.

For anyone arriving from a search box: the Sky Savings Rate is a governance-set yield accessed by supplying USDS and holding sUSDS. SKY governance votes on it, calibrated against Sky Protocol revenue, which is generated by the Sky Agent Network, an independent set of capital allocators that pay the protocol to access USDS liquidity under governance-set risk parameters.

A governance-set rate is not automatically better than a market-set one. It is differently exposed. A market rate carries market risk. A governance-set rate carries policy risk.

Sky showed this in 2026, when governance took the Sky Savings Rate from 4.75% toward the mid-threes to direct more Protocol Revenue into reserves. If you held sUSDS for the number, that happened to you. If you held it because you wanted a rate a solvent counterparty could keep paying, that was the point.

Question two: where does stablecoin yield come from?

Yield is never manufactured. It is transferred. Find the payer.

There are five in the whole market: borrowers paying for leverage, the US Treasury paying a coupon, long perpetual traders paying funding, a protocol redistributing its own revenue, or an issuer’s marketing budget. If you cannot find the first four, assume the fifth.

Sky’s model is the fourth. Sky Protocol makes USDS liquidity available to the Sky Agent Network. Independent allocators, currently including Spark, Grove, Obex and Osero, compete to deploy it across governance-approved strategies and pay the protocol for access. Those payments form Protocol Revenue.

Much published research is wrong here. Well-ranked explainers claim the Sky Savings Rate is funded by Treasury bills plus a Spark borrow rate, or that it tracks Fed funds with a thirty to sixty day lag. Both omit the Agent Network, which is the whole mechanism.

Believe it is a T-bill passthrough and you will expect it to move when the Fed moves.

Question three: what breaks it?

Everyone searching for the safest stablecoin yield is really asking this, and most answers they find are about audits. Audits are necessary. They are not where the money has gone.

TerraUSD went to zero in 2022 because its peg was an assumption, not a reserve. USDC touched $0.87 during the Silicon Valley Bank weekend in 2023 because the code was fine and the bank was not. Stream Finance froze because a manager lost money offchain. The dominant root cause of curator losses through 2025 and 2026 has been oracle misconfiguration: an audited contract reporting an audited price for an asset that had already collapsed.

So the useful version is not “has this been audited.” It is “how many independent things can break this, and can I name them all.”

Count them. A tokenised T-bill has three. A lending position, four. A curated vault, five, the fifth being curator discretion, which almost nobody prices. A basis-trade synthetic, six, because the hedge lives on centralised exchanges. sUSDS has three: contract surface, USDS collateral and peg behaviour under stress, and governance.

In August 2025, S&P Global assigned Sky Protocol a B- issuer credit rating, the first issued to a DeFi protocol. B- is speculative grade. The USDS stablecoin stability assessment was a 4, described as constrained, citing depositor concentration, governance concentration, and a risk-adjusted capital ratio below 0.5%.

Uncomfortable to publish, and the most useful paragraph here: most products you are comparing against have never been rated at all. Absence of a bad rating is not the presence of a good one.

Question four: can you leave?

Exit terms are the least-read section of any yield product, and the costliest to learn late.

• Contractual and immediate. sUSDS has no lockup, cooldown, queue or fee. One transaction.

• Contractual and delayed. sUSDe carries a seven-day cooldown, added in 2024 to manage redemption queues. Honest design. Seven days during which the thing you feared continues.

• Conditional. Most lending positions and curated vaults are instant until utilisation spikes. You find out the real number on the day you ask.

• Locked by design. Fixed Yield on Sky.money, using Pendle principal tokens, holds a rate to maturity. Not a defect. That is the trade.

Test the exit at your size, not the minimum.

Question five: can you verify it?

This is the only question that keeps working after you supply. One through four are diligence. Five is monitoring.

Live and onchain beats a periodic attestation, which beats reporting on a cadence the publisher chose, which beats a marketing page. A marketing page is not verification. It is a claim.

Sky Protocol publishes its balance sheet and income data live at financial.skyeco.com: Protocol Collateral, Sky Reserves, Gross and Net Protocol Revenue, per-agent deployment. It has posted five consecutive quarters of positive Net Protocol Surplus, with Q2 2026 Gross Protocol Revenue of $107.35 million against $97.15 million a year earlier.

Confirm one number from any marketing page using a source the marketer does not control. If you cannot, treat the page as a claim.

What the framework says about sUSDS

Roughly 3.65% on around $6.2 billion of supply in August 2026 is a governance-set rate on a fully liquid position with a short failure list.

sUSDe at a 9.4% seven-day average is a funding-market rate with six failure surfaces and a cooldown. Both are rational. They answer different questions.

If your question is “what is the highest number,” this framework will not lead you to sUSDS. If it is “what is the smallest set of things I need to be right about,” it might.

The two-minute version

1. Who sets the rate? Name the person, committee, market or formula.

2. What funds it? Name the payer, and what would make them stop.

3. What breaks it? Write the list. Size inversely to its length.

4. Can you leave? Find the cooldown, the maturity, the utilisation condition.

5. Can you verify it? Confirm one number from a source the marketer does not control.

The people who supplied into xUSD were not reckless. They saw a number that beat the alternatives. What they lacked was the habit of pulling the number apart before the money moved.

That habit costs ten minutes and some ego, because sometimes you run it on something you hold and dislike the result.

Run it on sUSDS. Run it on whatever you hold instead. The framework does not care which wins.

Access the Sky Savings Rate

sUSDS is the yield-generating stablecoin of Sky Protocol. Supply USDS through Sky.money to access the Sky Savings Rate, which is governance-set, fully liquid, and verifiable onchain at financial.skyeco.com. Rates are variable and set by SKY governance.

Sky.money is a non-custodial interface operated by Skybase International, an independent Sky Agent with no control over Sky Protocol smart contracts or ecosystem governance. Skybase does not set, control or guarantee the Sky Savings Rate or any other protocol parameter. Nothing here is investment, legal or tax advice. All figures are dated and subject to change; verify current data at financial.skyeco.com.

It doesn’t care about your brand story. It doesn’t watch your video. It reads your data, checks what others say about you, confirms your price and stock, and makes a call.

That salesperson is here. Gartner projects that by 2028, roughly 90% of B2B purchases - over $15 trillion in spend - will be intermediated by AI agents.

The question is no longer whether machines will buy. It’s whether they can read you well enough to buy from you.

There’s a phrase I keep coming back to: machine-readable or invisible. In an agentic market, those are the two states a brand can be in. If an agent can parse your facts, verify your trust signals, and act on your price and availability, you’re in the running. If it can’t, you don’t get a low score - you get no score. You simply aren’t there.

This issue is about the mechanics: how agents actually choose which brand to buy, what makes a brand visible to them, and the concrete steps to become machine-readable before your competitors do.

The 30-second version

• What changed: Agentic commerce is real infrastructure now. Open standards let agents authorize, check out, and settle payments - Google’s AP2 (60+ partners), OpenAI and Stripe’s Agentic Commerce Protocol, and Coinbase’s x402, which has already processed over 100 million transactions.

• Why it matters: Buyers already behave like machines. Forrester found 94% of B2B buyers use generative AI in purchasing; Gartner found 67% prefer a rep-free buying experience. The agent is doing the shortlisting - and it reads data, not decks.

• The takeaway: Completeness wins. Near-complete, structured product data (“golden records”) can make a brand 3–4x more likely to be recommended in AI shopping. Machine-readability is a growth lever now, not an IT chore.

In an agentic market there are only two kinds of brand: machine-readable, and invisible.

How do AI agents choose which brand to buy?

AI agents choose which brand to buy by reading structured information and third-party signals, then matching them against the buyer’s stated criteria.

An agent typically checks four things in sequence: your structured data and product feeds, third-party consensus about you (reviews, forums, listicles, knowledge bases), your real-time price and availability, and your trust signals (ratings, return policies, verifiable claims).

Brands that satisfy all four enter the shortlist; brands that miss one tend to drop out.

What an AI agent reads before it buys - miss one layer and you fall out of the trusted set.

Crucially, the agent doesn’t weigh your marketing. It weighs your legibility and your reputation.

Earned signals - what others say about you in machine-readable places - are several times more correlated with AI visibility than owned content, and structured knowledge like Wikidata is treated as near-axiomatic.

If the facts don’t line up, the agent defaults to the brand whose facts do.

What makes a brand visible to AI agents?

A brand is visible to AI agents when its key facts exist in a structured, consistent, machine-readable form across the places agents look.

Visibility comes from four things working together: clean structured data (schema markup, product feeds, an llms.txt file); consistent presence in third-party sources agents cite; accurate, real-time price and availability; and strong, verifiable trust signals.

Miss any one and you become harder to select - or invisible.

This is where completeness pays. Google’s AI shopping data shows that catalogs with near-complete attributes - the so-called golden record, around 99.9% attribute completion - can be 3 to 4 times more likely to surface in AI recommendations than sparse ones.

The Golden Record effect: near-complete data drives AI recommendation visibility. Source: Google AI shopping data.

How do I make my brand machine-readable?

To make your brand machine-readable, express your key facts in formats machines can parse and trust, everywhere they look.

In practice that means: add structured data (Organization, Product, FAQ, and Review schema) to your pages; publish and maintain accurate product feeds; add an llms.txt file summarizing what your brand is and offers; keep price and availability current and accessible; and build consistent third-party presence - reviews, comparison articles, and knowledge-base entries - so the agent finds corroboration, not contradiction.

The agentic payments stack: authorize, check out, settle.

It helps to know how the transaction closes, too. Three open standards now let agents pay: AP2 handles authorization, the Agentic Commerce Protocol handles checkout, and x402 handles settlement.

The rails are live. The brands preparing their data for them now are the ones agents will transact with first.

Are buyers really ready for this?

Yes - buyer behavior is already agent-shaped. The demand side moved before most brands did.

B2B buyers already behave like agents. Sources: Forrester, Gartner, Digital Sales Institute, BCG.

Forrester found 94% of B2B buyers now use generative AI in the purchasing process. Gartner found 67% prefer a rep-free buying experience, and separate research found 73% are willing to place orders over $50,000 self-serve.

Yet BCG found only 18% of B2B companies rate their own AI-commerce maturity as advanced. The gap between how buyers act and how brands are prepared is the opportunity - and it’s closing.

$15 trillion in B2B spend expected to be intermediated by AI agents by 2028. Source: Gartner.

RUN THIS YOURSELF (20 MINUTES)

Pick one flagship product. Then play the agent.

(1) Open your product page’s source and check for structured data - is there Product schema with price, availability, and reviews? (Use Google’s Rich Results Test.)

(2) Ask an AI agent to compare your product to two competitors “using only publicly available data.” Note every fact it gets wrong or can’t find.

(3) Search for your product on Reddit and in comparison articles. Is there anything for an agent to cite?

Each blank you find is a place you’re currently invisible. Fixing them is machine-readability, one field at a time.

Frequently asked questions

Q. What is a machine-readable brand?

A. A machine-readable brand is one whose key facts - what it sells, at what price, with what availability, and with what proof - are published in structured, consistent formats that AI systems can parse and trust.

It’s the prerequisite for being recommended by AI agents.

Q. What is a golden record in AI commerce?

A. A golden record is a complete, accurate, single source of truth for a product’s attributes.

Near-complete golden records (around 99.9% attribute completion) can be several times more likely to appear in AI-generated shopping recommendations, because agents favor the data they can fully verify.

What are AP2, ACP, and x402?

They’re the open standards behind agentic payments. AP2 (Google) handles authorization - proving an agent is allowed to buy.

ACP, the Agentic Commerce Protocol (OpenAI and Stripe), handles checkout. x402 (Coinbase, with Cloudflare) handles settlement and has already processed over 100 million transactions.

Go deeper: The full playbook - the four-layer model agents use, the schema and feed checklist, and how to prepare for AP2, ACP, and x402 - is on Medium.

→ Read the full guide: Machine-Readable or Invisible: Why Brands Must Be Legible to AI Agents

FOLLOW US

LinkedIn - Utkarsh Tyagi

X (formerly Twitter) - @agenticguy

That’s the first three issues. Next time we go deeper into building your Share of Model - the metric that ties this whole series together.

— Utkarsh

I asked him to do one thing: open Perplexity and ask the exact question his customers ask. He sent a screenshot back twenty minutes later. His #1-ranked page wasn’t cited anywhere in the answer. Three other sources were — a Reddit thread, a comparison article on a site he’d never heard of, and his competitor’s docs.

He hadn’t lost his ranking. He’d lost something he didn’t know existed: the citation.

“Is SEO dead?” is the wrong question. SEO is very much alive — Google still drives enormous traffic, and strong SEO still correlates with getting cited by AI. But a quiet decoupling has happened underneath it. Ranking high and getting cited by an AI are now two different games, won in two different ways. If you only play the first, you’re leaving the second — the faster-growing one — on the table.

The 30-second version

• What changed: The overlap between the pages Google ranks first and the sources AI engines actually cite has fallen from around 70% to under 20% in three years (Brandlight). Ranking and citation have come apart.

• Why it matters: Ahrefs found fewer than 9% of the sources ChatGPT and Gemini cite come from Google’s top 10 — meaning ~91% come from somewhere else. Optimizing only for rank leaves most AI citations unclaimed.

• The takeaway: Add a new layer on top of SEO — Answer Engine Optimization (AEO). Same foundation, different objective: not to rank a page, but to be the source a model selects.

SEO earns you a position. AEO earns you a citation. In 2026, the second one is the one that gets recommended.

Is SEO dead in 2026?

No — SEO is not dead in 2026. Search still sends billions of visits, and pages with strong SEO are more likely, not less, to be cited by AI engines. What has changed is that ranking is no longer sufficient on its own: being the #1 result and being the source an AI cites are now separate outcomes that require overlapping but distinct optimization.

The confusion comes from conflating two trends. Zero-click search is rising — Google’s AI Overviews reach over 2 billion users a month, and a majority of searches now end without a click. At the same time, AI-referred visitors convert at several times the rate of ordinary search traffic. So there’s less traffic, but the traffic that arrives via AI is more valuable. SEO isn’t dying; it’s splitting into “rank for humans” and “get cited by machines.”

What is the difference between SEO and AEO?

The difference between SEO and AEO is the target. SEO (Search Engine Optimization) optimizes a page to rank in a list of links for human clicks. AEO (Answer Engine Optimization) optimizes content and data to be selected and cited as a source inside an AI-generated answer. SEO competes for position; AEO competes for citation.

SEO and AEO side by side — a new layer, not a replacement.

As Kelsey Voss put it, SEO is about ranking pages for clicks, while GEO/AEO is about being selected as a source. And as Lily Ray notes, this isn’t a replacement — it’s an additional layer. The mechanics differ: SEO leans on keywords, backlinks, and technical performance; AEO leans on clear definitions, cited statistics, structured data, and third-party consensus.

Overlap between Google’s top links and AI citations, 2023–2026. Source: Brandlight.

Why doesn’t ranking get you cited by AI?

Ranking doesn’t automatically get you cited because AI engines don’t simply read Google’s top 10 and summarize it. They run their own retrieval — often expanding a single question into several sub-queries (a process called query fan-out), pulling sources for each, and then synthesizing an answer that names only one or two brands. A page can rank #1 and never enter that retrieved set.

Query fan-out: one question becomes several, and only a few sources make the answer.

There are three common reasons a top-ranked page gets skipped: it’s thin or persuasive rather than factual (nothing quotable to lift); it lacks cited statistics and clear definitions (the elements models prefer); or the citation is being won elsewhere — on Reddit, in a well-sourced listicle, in documentation. BrightEdge found 42% of AI citations come from outside the organic top 10; Ahrefs put the share of AI citations from Google’s top 10 at under 9%.

Ranking strength vs AI citation likelihood — the two axes that now matter.

What actually makes a page get cited?

Pages get cited when they make an AI’s job easy: a clear answer near the top, relevant statistics with sources, plain definitions, and structure a model can parse. A Princeton study on generative engine optimization found that adding cited statistics, quotations, and authoritative sources increased a page’s likelihood of being cited by up to 52% — one of the largest single levers measured.

Adding cited statistics is among the biggest levers for AI citation. Source: Princeton GEO study.

RUN THIS YOURSELF (15 MINUTES)

Pick your most important keyword — the one you rank well for. Now run the underlying question through Perplexity and ChatGPT with web search on.

Write down: (1) Is your page cited? (2) Which sources are cited instead? (3) What do those sources have that yours doesn’t — a stat, a definition, a table, a clear answer up top?

That gap list is your AEO backlog. Nine times out of ten, the winning sources aren’t better written. They’re more liftable.

Frequently asked questions

Q. Should I stop doing SEO and switch to AEO?

A. No. AEO builds on SEO — strong technical SEO and authority still help you get cited. Treat AEO as an added layer: keep the SEO foundation, then optimize the same pages to be answer-ready with clear definitions, sourced stats, and structured data.

Q. Is AEO the same as GEO?

A. They’re used almost interchangeably. AEO (Answer Engine Optimization) emphasizes being cited in answer engines; GEO (Generative Engine Optimization) emphasizes visibility inside generative models. In practice the tactics overlap heavily — both are about being selected as a source rather than ranking a link.

Q. How fast do AI citations change?

A. Quickly. Studies have found that 40–60% of the sources AI engines cite can change month to month, so AEO is an ongoing practice, not a one-time fix. Freshness and consistent third-party presence matter more than in classic SEO.

Go deeper: I broke this down in full on Medium — the data behind the decoupling, the exact page-level changes that win citations, and a checklist to turn ranking pages into cited ones.

→ Read the full guide: SEO Is Not Dead, But Ranking No Longer Means Being Cited

FOLLOW US

LinkedIn — Utkarsh Tyagi

X (formerly Twitter) — @agenticguy

Next issue: what makes a brand visible to AI agents — and why “machine-readable or invisible” is the choice every brand now faces.

— Utkarsh

Subscribe now

Last week I watched someone buy a $4,000 espresso machine without reading a single review.

She typed one sentence into an AI assistant - “find me the best prosumer espresso machine under $4k for a small cafe, and order it” - and then closed the laptop. The agent read the specs, weighed the trade-offs, checked stock and delivery, and placed the order. She never saw a search results page. She never clicked an ad. She never visited a brand’s website.

Three brands were in the running. Two of them lost the sale before a human ever looked at them. Not because their product was worse - because the machine couldn’t read them clearly, and the one it could read, it trusted.

That is agentic marketing in one scene: the moment the buyer stops being the person and starts being the person’s software.

For twenty years, marketing has been a fight for human attention - the click, the scroll, the impression. That fight isn’t over, but a second one has opened up beside it, and it’s growing far faster: the fight to be understood, trusted, and recommended by the AI agents that increasingly sit between your brand and your customer.

This is the first issue in a series on how to win that fight. Today: what agentic marketing is, why 2026 is the year it stopped being theoretical, and what it changes for you starting now.

The 30-second version

• What changed: Buyers are delegating research and purchases to AI agents. Gartner expects 40% of enterprise apps to ship task-specific agents by the end of 2026, up from under 5% a year ago.

• Why it matters: Your next “customer” is often a model deciding which brands to shortlist. If the model can’t read your facts or doesn’t trust your signals, you’re invisible — no matter how good your product is.

• The takeaway: Stop optimizing only to persuade people. Start optimizing to be cited by machines. The new scoreboard isn’t traffic - it’s Share of Model.

Agentic marketing is the practice of making your brand legible, trustworthy, and selectable to autonomous AI agents - so that when a machine decides, it decides for you.

What is agentic marketing?

Agentic marketing is a marketing approach designed for autonomous AI agents rather than only human audiences. It optimizes a brand’s content, data, and reputation so that AI systems - the assistants and agents that research, compare, and increasingly purchase on a person’s behalf - can accurately understand the brand and recommend it.

Where traditional digital marketing tries to win a human’s click, agentic marketing tries to win a machine’s citation. The “agent” in question perceives a goal, plans the steps, evaluates options against criteria, and acts - booking, buying, or shortlisting — often with little or no human review of the individual choices.

It sits alongside a fast-growing vocabulary: Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO) describe the tactics; zero-click describes the behavior; Share of Model describes the scoreboard. The common thread is that large language models - ChatGPT, Perplexity, Claude, Gemini, Grok, Copilot - are now a primary interface between brands and buyers.

How is AI changing marketing in 2026?

AI is changing marketing in 2026 by moving the point of decision away from the human and toward the agent. Instead of browsing and choosing, people increasingly delegate the browsing and the choosing to AI - which means brands are now evaluated by software before a person ever sees them.

The shift is not gradual. Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from fewer than 5% in 2025 — one of the steepest adoption curves since mobile. Worldwide AI spending is on track to reach $2.52 trillion in 2026, and Gartner expects 60% of brands to use agentic AI for one-to-one customer interactions by 2028.

Enterprise adoption of task-specific AI agents, 2024–2026e. Source: Gartner enterprise AI forecasts.

For marketers, three things change at once: discovery moves inside AI answers instead of search pages; evaluation is done by a model reading your data, not a person reading your copy; and the winning metric shifts from clicks and traffic to how often you’re named in the answer.

What does agentic AI mean for brands?

For brands, agentic AI means the audience has split in two: the human you’ve always marketed to, and the machine that now decides on the human’s behalf. Winning requires being persuasive to people and legible to models at the same time - and the second audience is the one most brands are ignoring.

The shift from marketing to people to marketing to models.

The practical consequence is a new scoreboard. In a search world you measured rank and traffic. In an agentic world you measure Share of Model: how often, and how favorably, your brand is named inside AI answers for the questions that matter to your category.

Share of Model: how often each brand is named inside AI answers for a category (illustrative).

A brand can lose 90% of its search clicks to zero-click answers and still win - if it’s the brand the answer recommends.

RUN THIS YOURSELF (10 MINUTES)

Open ChatGPT, Perplexity, Gemini, and Claude. Ask each the exact question your best customer would ask before buying in your category - for example, “what’s the best [your product] for [use case], and why?” Do it in a fresh session with no personalization.

Now read the answers like a scoreboard. Are you named? Are your competitors named more often? Is the description of you accurate? Is anything wrong or outdated?

Whatever those four answers say is your real starting position in the agentic market. Most brands are shocked - not because they rank badly, but because they aren’t in the conversation at all.

The new buying path (and where brands fall out)

Here’s the loop that’s replacing the funnel. A human states an intent once. An AI agent perceives that intent, plans the steps, and assembles a shortlist - a trusted set of a few brands. Then it acts: it picks one and skips the rest.

The agentic buying loop: intent stated once, the agent does the rest.

Every stage is a place you can win or disappear. You fall out at “perceive” if the agent can’t parse your product data. You fall out at “trusted set” if third-party signals - reviews, forums, structured facts - don’t back you up. And you fall out at “act” if your price, availability, or claims aren’t machine-readable in the moment of decision.

Notice what’s missing from this loop: your homepage hero, your brand video, your clever tagline. The agent doesn’t watch, scroll, or feel. It reads. Agentic marketing is the discipline of making sure that when it reads, it reads you clearly — and chooses you.

Frequently asked questions

Q. Is agentic marketing the same as AEO or GEO?

A. Not quite. AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization) are the tactics - structuring content and data so AI systems cite you. Agentic marketing is the broader strategy that includes those tactics but also covers reputation, third-party presence, and being selectable at the moment an agent acts, not just cited in an answer.

Q. Does agentic marketing replace traditional marketing?

A. No. It’s an additional layer. Humans still discover, feel, and decide - and human-facing brand, story, and design still matter. Agentic marketing adds a second audience, the machine, that increasingly gates access to the first. The brands that win in 2026 do both.

Q. How do I measure agentic marketing?

A. With Share of Model - how often your brand is named, and how favorably, across AI answers for your category’s key questions. You can track it manually by testing prompts across ChatGPT, Perplexity, Gemini, and Claude, or with tools like Profound, Semrush, and Otterly that measure AI visibility at scale.

Go deeper: I wrote a full breakdown of the 2026 definition - the frameworks, the data, and a step-by-step way to audit your own brand’s agentic visibility — on Medium.

→ Read the full guide: What Is Agentic Marketing? The 2026 Definition for Brands

FOLLOW US

LinkedIn - Utkarsh Tyagi

X (formerly Twitter) - @agenticguy

Next issue: why SEO isn’t dead - but ranking no longer means being cited. See you then.

— Utkarsh