See Payani AI in action. Watch the demo

Payani AI
The blog
Insights

What AI Assistants Cite When They Recommend a Vendor

PAYANI AINovember 2, 20269 min read

What AI Assistants Cite When They Recommend a Vendor
ShareLinkedInX

Your category gets recommended every day by ChatGPT, Gemini, Perplexity and Google AI Mode, and your company is not in the answer. The reason is not that your product is worse. It is that an assistant does not pick the best vendor, it assembles an answer from sources it already trusts, and nothing it trusts says anything specific and checkable about you.

That is an uncomfortable diagnosis, because it means the fix is not more content. It is a different kind of content, published in a different shape, in more than one place.

Ranking and being cited are two different jobs

Classical search returns a ranked index. You compete for a position, and the position is the prize. An answer engine does something else. It retrieves passages from sources it has reason to trust, synthesises them into a claim, and names whoever supplied the fact it used. A 2026 preprint on measuring brand visibility across AI search engines describes the same shift in terms, from search engine optimisation to generative and answer engine optimisation. The practical split is this: ranking on a search engine, being cited when a model synthesises an answer, and whether your business is represented inside the model's retrieval at all.

So page one is not citation. You can rank for "best [category] software" and never be named, because the sentence the model needed was not on your page in a liftable form, and nobody else stated it about you either.

The second uncomfortable part is where the retrieval happens. Semrush analysed 325,000 unique prompts submitted between January and February 2026 across ChatGPT Search, Google AI Mode and Perplexity, across 12 industry categories, and found LinkedIn appeared in 11% of AI responses on average, the second most cited domain in that dataset, with 89,000 LinkedIn URLs examined. That is a vendor study and should be read as one, but the direction matters: a meaningful share of what gets said about your company is said somewhere that is not your website. Vendor analyses of citation sources also put community platforms near the top of the list, with Reddit the single most-cited domain across the major AI engines. Evertune's analysis found even the most-cited domain on any platform rarely exceeds 5 percent of total citations, with the other 95 percent spread across thousands of domains. Your domain is one input among several.

The four things that get a vendor named

Across the evidence, the pattern of what earns a citation is narrow and consistent.

1. A specific claim. Not a capability, a fact. "Migrates from your existing CRM in a day" is a claim. "Powerful, flexible CRM" is a mood. Models cite things that can be stated as an answer.

2. A source that states it plainly. The claim has to appear as a clean sentence a model can lift without reconstructing it from three paragraphs of positioning. The GEO study from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi tested a set of content changes and found that adding relevant statistics, quotations and citations boosted source visibility in generative engine responses by up to 40%, with the effect varying by query domain.

3. Corroboration. One page saying it is an assertion. The same fact appearing on your site, in a founder post, in a customer's own words, in a directory listing and in a third-party writeup is a pattern the retrieval layer can trust.

4. Machine-readable structure. Clear headings, direct question-and-answer formatting, consistent entity naming, clean markup. Treat this as hygiene that makes your claims legible, not as a lever with a number attached to it.

The more useful half of that study is what did not work. The tactic that failed was keyword stuffing, which adds query terms to the page. The authoritative rewrite did help, but mainly on debate and history style queries, not on factual ones, where citations won. So the models are not counting your keywords, and persuasion does not beat specificity when a buyer asks a question of fact.

The lift test

Here is the test we apply to every article before it goes out. Take any sentence on the page and ask: if an assistant were answering a buyer's question right now, could it quote this sentence verbatim as the answer, and would the quote still be true out of context?

Most marketing copy fails immediately. "We help growth teams move faster" answers nothing. "A workflow trigger model that fires when an event happens, only if your conditions are met, described in plain language" is quotable, specific and checkable. One of those sentences can be cited. The other cannot be used for anything.

Run the test across your product pages. You will usually find that the only liftable sentences on the whole site are in the pricing table and the footer.

Authority also behaves differently than most advice assumes. An SE Ranking analysis of 129,000 unique domains, as reported by Contently, found .gov and .edu domains underperformed commercial domains on general queries, averaging 3.2 citations versus 4.0 for commercial domains, while outperforming on health, science, finance and policy topics. That split varies by topic, so advice drawn from health, science, finance or policy queries may not carry over to a commercial "who should I use for X" question.

The audit to run this week

This takes a couple of hours and it is the only way to know where you actually stand.

  1. Write the ten buying questions in your category. The real ones, in buyer language. "Best CRM for a team switching off a fragmented stack." "What tool connects Google Ads spend to closed revenue." Not your keyword list.
  2. Ask three assistants each question. ChatGPT, Gemini and Perplexity is a reasonable spread. Use a fresh session each time.
  3. Record three columns. Were you named, who was named instead, and what source was cited for them. The third column is the one that pays.
  4. Read the cited sources. You are looking for the sentence that got quoted. Nine times out of ten it is a plain factual statement, often on a page nobody would call a marketing asset.
  5. List the gaps. Every question where you were absent is a claim you have never published in a liftable form.

That gap list is your content plan for the quarter. Not topics, claims.

What to publish, and what to stay away from

Publish narrow first-party facts under your own name. How the thing actually works. What it costs. What it does not do. Named customer outcomes with the number in the sentence. Comparison pages that state the real tradeoff rather than winning every row. Founder and operator commentary on your own surfaces, since professional profile content is clearly part of the retrieval pool in that Semrush dataset.

What to stay away from: manufacturing third-party enthusiasm. Community platforms have self-promotion rules, and disclosure obligations apply to incentivised mentions and endorsements. Seeding praise you did not earn is a platform violation and a legal exposure, and the specifics vary by platform and jurisdiction, so get your own counsel to review anything close to the line. The operator version is simple. Publish checkable facts under your own name, and let other people repeat them because they are true.

Telling improvement from noise

Assistants give different answers to the same prompt on different days. A single check proves nothing, and most teams declare victory or defeat on exactly that.

Fix the prompt set at ten to twenty questions. Run each one three times, in fresh sessions, on a fixed cadence. Record presence rate, not a yes or no. Read it quarterly, because content needs time to be crawled, corroborated and retrieved.

Then connect it to revenue, which is the part almost nobody does. Citation count is a vanity metric. The real question is whether assistant-referred visitors land on your site, convert on a form, and become pipeline. That is a tracking and attribution question, and it is answerable: visitor tracking that identifies the companies arriving, forms that capture them, and reporting that follows the same record through to a closed deal.

This is why AI search visibility sits inside the Payani AI platform rather than beside it. You score how you show up today, publish the researched articles that close the gaps, and watch the visitors those answers send you turn into contacts, deals and revenue in the same system. One loop, one set of records, no guessing at the middle.

Frequently asked questions

Why does ChatGPT recommend my competitors and not us?

Because an assistant builds an answer by retrieving passages from sources it already trusts, rather than ranking an index. If no trusted source states a specific, checkable fact about your company in a form that can be quoted, there is nothing for the model to cite. Your competitors are usually not better, they are just more quotable in more places.

Does ranking on page one of Google get me cited by AI assistants?

Not automatically. Ranking and citation are separate outcomes, and research on generative engines treats them as different things entirely. You can hold a top position and still never be named, because the model needed a liftable factual sentence and your page offered positioning instead.

What actually increases the chance of being cited?

The GEO study from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi found that adding relevant statistics, quotations and citations boosted source visibility in generative engine responses by up to 40%, with the effect varying by query domain. The same work tested an authoritative-sounding rewrite and keyword stuffing, which are the two things most teams still default to. Specificity beats persuasion.

How do I know if my AI search visibility is actually improving?

Fix a prompt set of ten to twenty buying questions, run each one three times in fresh sessions on a regular cadence, and record a presence rate rather than a yes or no. Assistants vary their answers day to day, so a single check tells you nothing. Read the trend quarterly, and track whether assistant-referred visitors convert into pipeline rather than counting citations.

Should we be posting in communities and on LinkedIn to get cited?

A share of citation happens off your own domain. Semrush's analysis of 325,000 prompts across ChatGPT Search, Google AI Mode and Perplexity found LinkedIn appeared in 11% of AI responses on average, the second most cited domain in that dataset. Publish under your own name and state real facts, but do not manufacture third-party enthusiasm, since self-promotion rules and endorsement disclosure obligations apply and the specifics warrant legal review.

ShareLinkedInX

Newsletter

Get growth intelligence in your inbox.

Sharp notes on growing with AI and what actually moves revenue. No fluff, unsubscribe anytime.