The Machines Check Your References

On August 6 we pointed our own scanning pipeline at our own products. Thirty-eight buyer prompts, repeated runs across ChatGPT, Perplexity, and Gemini, 209 scored samples, every raw answer kept on disk with a timestamp.

One of the prompts was “Suno alternative that doesn’t take rights to my songs.” Rights are our entire pitch. We were absent. One captured run named beatoven.ai, twoshot.app, and soundraw.io instead. Another named Suno itself, which is the product the buyer was trying to leave.

We published the losses, in full tables, before charging anyone for the work. Nineteen of the 38 prompts returned answers naming competitors only. A Suede product appeared in 30.6% of the 209 samples. The teardown is public at scan.suedeai.ai/teardown/suede, and this essay is what running it taught me about how visibility actually works now, and why we rebuilt our whole practice around it at seo.suedeai.ai.

The vocabulary in this field is a mess, so here are the working definitions:

• SEO (search engine optimization): earning placement in a ranked list of links.

• AEO (answer engine optimization): making your facts citable inside the AI-generated answers that ChatGPT, Perplexity, Gemini, and Google’s AI results return.

• GEO (generative engine optimization): the academic name for roughly the same discipline as AEO. Different communities, same job: influence what generative engines say and cite.

• Digital PR: earning coverage from publications machines already trust.

• Reputation management: keeping the third-party record about you accurate and consistent, everywhere machines look.

In the link era you could run these as separate budgets with separate agencies, and the seams didn’t show. A ranking is forgiving. Ten blue links means ten winners per query, and a mediocre reputation could still hold slot seven.

An answer is not forgiving. When a buyer asks an assistant “what’s the best invoicing tool for freelancers, which one should I actually use,” they get one synthesized paragraph and a handful of citations. There is no slot seven. You are in the answer or you are not, and the engine decided that before the buyer finished typing. The first thing a buyer reads about you is now written by a machine, and it decides who gets the call before your site is ever opened.

That is why the five disciplines collapsed. The engine does not check your rankings, your press clips, and your reviews separately. It resolves one question: do I know who this company is, and do I trust what I know enough to repeat it? Every prong of visibility work now feeds that single decision. We run the practice as one retainer across four fronts, GEO, PR, reputation, and classic search, because splitting them is pretending the machine grades them separately. It doesn’t.

The mechanism worth understanding: a search engine ranked documents. An answer engine resolves entities.

When an assistant composes an answer, it retrieves candidate sources, extracts claims, and has to decide which entities it is confident about: this company exists, it does this thing, for these people, and these sources agree. Confidence comes from corroboration. A fact that appears once, on your own site, in your own words, is an assertion. The same fact repeated by sources the engine already trusts is a record.

Our own teardown shows the pattern plainly. Prompts about the founder were cited in 93.1% of samples. Prompts about the company, 88.9%. Prompts about individual products, far worse, down to absent. The person and the company have years of corroborated record: repositories, bylines, registries, published work that all say the same thing. The newer products had months. The engines cited exactly in proportion to how much third-party record existed. Nothing about that is mysterious. The machines check your references, and the young products hadn’t accumulated any yet.

Look at who got cited instead of us: copyright.gov, Songtrust, Bandcamp, GitHub, established products with a decade of coverage. The engines reached for entities with long, boring, consistent third-party records. Losing to copyright.gov on a song-ownership prompt is not a ranking failure. It is the graph telling you who it trusts.

A knowledge graph is a database of entities, people, companies, and products, plus the relationships between them, assembled from sources the builder trusts: structured data on your own site, established publications, directories with editorial standards, public registries, code repositories, filings. Google’s Knowledge Graph is the famous one. Every AI lab now runs some blend of the same idea through training data and retrieval.

The unit is the entity, not the page. That inverts most content strategy. A thousand blog posts that each describe your company slightly differently do not build an entity. They build noise. What builds an entity is the same canonical facts, in the same words, confirmed across surfaces that don’t share an owner: your site says it, your schema markup says it, your GitHub says it, your registry listings say it, and then, decisively, somebody who is not you says it.

Contradictions work in reverse. When records disagree, name variants, stale descriptions, a founder bio that differs across three profiles, the engine doesn’t average them. It loses confidence and cites someone cleaner.

Here is the part most visibility advice gets wrong: corroboration is weighted, not counted.

Engines do not tally mentions. A citation from an entity the graph already trusts, an established publication, a government registry, an open-source project with a maintainer who says no, carries the trust of that entity’s own record with it. Fifty listings in paid directories that accept anyone corroborate nothing, because the corroborating entities themselves have no record worth trusting. You cannot vouch for yourself, and neither can a source that vouches for everyone.

This is why the junk shortcuts backfire. Press-release wire spam, purchased backlink farms, fake review batches, and AI-generated directory sites don’t add weak positive signal. They attach your entity to low-trust nodes and manufacture contradictions between records, and contradictions are precisely what drops an entity’s confidence. A thin graph made of five reputable records beats a thick graph made of five hundred junk ones, because the machine can confirm the thin one.

The old-world version of this rule was “you are the company you keep.” The machine-readable version is stricter, because the machine never forgets an edge and never got charmed at a dinner.

We did not learn this from a whitepaper. The GEO and reputation practice exists because the Suede estate needed it first: 30+ live sites and 8 iOS apps competing for AI answers, measured continuously with in-house tooling. What worked there is what we sell, and the record it produced is public on the practice page:

• Open source accepted by maintainers. 32 pull requests merged across 30 open-source projects as of August 2026, including Jest, Adobe React Spectrum, Backstage, and Hardhat. Each one is a third party with its own standing saying yes to the work. Every tile on the page links to the merged PR, and the GitHub search that proves the count is one anyone can run.

• A platform partner on the record. In partnership with Google Cloud since September 2025, with the Suede stack on the Vertex AI suite and the account team’s proposal to collaborate on a launch story in writing.

• Government registries. Patent 63/947,120 on file with the USPTO for the Suede provenance system.

• Editorial coverage. TechBullion in May 2026 and Programming Insider, both linked from the page rather than paraphrased.

• Institutions with admission standards. Stripe Startups in July 2026, and enrollment at Florida Atlantic University’s Entrepreneurship Institute.

Notice what every item has in common: none of it is self-published, and every claim links to a source a buyer or a machine can check. A merged PR is Adobe’s maintainers vouching. A patent filing is a federal register vouching that the application exists. A partnership quote in writing is Google’s account team vouching. That list is a knowledge graph under construction, edge by edge, each edge terminating at a node the engines already resolved and already trust.

And the vouching runs both ways, which is the part almost nobody prices in. Your associations are part of your record. The clients an agency signs, the partners a startup lists, the directories a company submits to, all of it becomes edges in the same graph. It is why we keep the client roster deliberately small and decline work we wouldn’t put our name against: every engagement ties into the same public entity record the practice runs on, and it has to hold up there.

One retainer, four fronts, one record:

1. GEO / AI answers. What ChatGPT, Perplexity, Gemini, and Google’s AI results say about you, captured with dates and screenshots, then repaired at the source pages the engines are reading. Our own repair pass included auditing 416 llms.txt links across 21 domains down to zero broken.

2. PR / press. The founder story, the newsroom, the pitch material, and the outreach behind coverage. When something lands it becomes more evidence on the record. No outlet’s decision is ever promised, because it can’t honestly be.

3. Reputation. The compounding layer: knowledge panels, entity graphs, third-party records, and monitoring of how machines describe you, because records rot quietly and engines re-read them constantly.

4. SEO / search. The classic work done properly: technical repair, content that answers what buyers actually ask, entity records that hold up, internal architecture engines can follow.

Month one always opens with the capture, pointed at your category. Not because measurement is the product, but because the misses are the roadmap. Our 19 invisible prompts told us exactly which entity records didn’t exist yet. A score tells you to worry. Evidence tells you what to build.

The work is unglamorous, which is a reasonable sign it’s real:

1. Write one canonical record. Legal name, one-sentence description, founder, category, links. Repeat it verbatim on every surface you control. Every variant you allow is a contradiction you paid for.

2. Ship the structured data. Organization, Person, and Product schema with sameAs links pointing at every real profile. This is you filling out your own graph paperwork so nobody has to infer it.

3. Check that the crawlers can even read you. GPTBot, PerplexityBot, ClaudeBot, and Google-Extended have to reach the pages before anything else matters. We publish a free live check at optimize.suedeai.ai.

4. Build citable pages. One page per real buyer question, answering it in a block a machine can lift whole: a definition, a number, a comparison, a date.

5. Earn coverage slowly, from real publications. One piece in an outlet with editorial standards outweighs a quarter of wire releases. Pitch the story that is true.

6. Claim the boring third-party records. Registries, app stores, GitHub, patent filings, institutions that say no to people. Unglamorous records are the load-bearing ones.

7. Monitor for drift. Re-run your buyer prompts on a schedule. Answers move between runs, and drift is invisible from inside your own building.

Nobody can promise you a citation. AI answers are non-deterministic; they change by engine, by phrasing, by account context, and by the hour. Every capture in our reports is labeled a point-in-time snapshot with the run count stated, because that is what it is. The work targets the durable inputs, schema, entity records, citable pages, earned corroboration, and then measures again to show the trend. Anyone who guarantees “we’ll make ChatGPT recommend you” is selling a control the engines do not offer them. Ask them for their raw captures and watch what happens.

What is AEO in one sentence?

Answer engine optimization is the practice of making your company’s facts consistent, machine-readable, and corroborated by trusted third parties, so AI assistants can cite you in their answers with confidence.

What’s the difference between AEO, GEO, and SEO?

SEO earns placement in a ranked list of links. AEO and GEO are near-synonyms for earning citation inside the single answer an AI engine writes; GEO is the term the academic literature uses, AEO the one practitioners use. The tactics overlap, but the scoreboard changes: instead of a rank position, you either appear in the answer or you do not.

What is a knowledge graph, practically?

The database of entities and relationships an engine trusts enough to answer from. You influence yours by making the same canonical facts appear in your structured data, your public records, and coverage from sources with their own standing.

How do AI engines decide which companies to cite?

Nobody outside the labs knows the exact weighting, and it changes. The observable pattern in captured answers is consistent: entities with long, consistent, third-party-corroborated records get named; entities that only describe themselves get skipped. Our own teardown showed a 93.1% citation rate for the founder’s well-corroborated record against near-invisibility for months-old products.

Can a small company do this without a PR budget?

Yes, up to a point. The canonical record, schema markup, crawler access, citable pages, and boring registries cost discipline, not money, and they are the floor everything else stands on. Earned press amplifies a consistent record; it cannot substitute for one.

The full method is published free, no email: “The Screenshot: Why AI Recommends Your Competitors, and How to Fix It,” all eleven chapters, plus the audit checklist, at seo.suedeai.ai. The teardown we ran on ourselves, losses included, is at scan.suedeai.ai/teardown/suede. Or run the cheapest version of the experiment tonight: type what you sell into ChatGPT the way a buyer would, and see whose references were in order.

I write about AI visibility, creator rights, and building Suede in the open. Subscribe for the next essay.