How AI engines decide which brands to cite

An AI engine doesn't decide to cite you. It decides to cite a passage, and your brand is a side effect of which passages made it into the context window.

That reframe is the whole article, but let's do the mechanics properly, because the details are where the actionable work lives.

What happens between the question and the answer?

Three stages, in order, and you can be eliminated at any of them.

Stage 1: Query fan-out. The engine takes the user's conversational question and rewrites it into several underlying search queries. Ask "what's the best AI visibility tracker for a small agency that needs Perplexity coverage" and the engine doesn't search that string. It fires something closer to ai visibility tracking tools, best geo tools for agencies, perplexity brand monitoring, maybe ai visibility tracker pricing. Each of those hits an index.

Stage 2: Retrieval. Each query returns candidates. The engine pulls a few dozen documents, chunks them into passages, and re-ranks those passages by relevance to the original question, not the rewritten one. What survives is maybe three to ten passages that go into the model's context window.

Stage 3: Grounding and generation. The model writes an answer using those passages as evidence, and attaches citations to the sources that support each claim. Modern engines run a verification pass here: a claim that isn't supported by a retrieved passage gets dropped or softened, because unsupported claims are how you get hallucination headlines.

Your brand gets named when it appears, consistently, across enough of the passages that reach stage 3. That's it. That's the mechanism.

Why does consensus beat one great page?

Because the model is trying not to be wrong, and one source saying something is weak evidence.

When a model generates "the leading options are A, B, and C," that's a factual claim about the world. The grounding step wants support for it. If four independent retrieved passages all list you among the leading options, that's strong support. If the only passage that names you is your own homepage saying you're the leading option, that's an advertisement, and the model treats it accordingly.

This is the part where GEO stops feeling like SEO. You can write the single best page on the internet about your category and still lose the answer, because the contest isn't "whose page is best." It's "what do the retrieved sources, collectively, say is true."

Practical version of that: for any buyer question you care about, pull the sources the engine actually cited. If five of the seven are third-party listicles and you're in none of them, you have an outreach problem, not a content problem, and publishing another blog post will not fix it. Getting into three of those five listicles will.

We see this constantly in scan data. Brands with strong domain authority and excellent on-site content get retrieved and then not mentioned, because every corroborating source in the retrieved set is a roundup they aren't in.

How does each engine source differently?

They differ enough that "optimizing for AI" as a single activity is a mistake. Here's what actually varies:

ChatGPT

ChatGPT answers from model knowledge by default, and searches the web only when it decides the question needs fresh information. That decision is invisible to you and it's the single biggest source of confusion when people spot-check their brand.

Two consequences:

1. For questions that feel timeless ("what's a good CRM for a small team"), you may be competing against the model's training data, which means the web as it looked months ago. You cannot fix that with a page published yesterday. What you can fix is what the web will look like the next time a snapshot is taken, and what the search index returns when a search does fire.
2. When it does search, it uses its own index via OAI-SearchBot, which is not Google's index. Ranking on Google helps you indirectly, at best.

ChatGPT also changed the economics of the mention recently. When OpenAI put clickable brand links into answers, Similarweb measured referral traffic rising 157.7% week over week. With 900M weekly active users, a mention here is the highest-value slot in the category.

Perplexity

Perplexity is the most transparent engine and the easiest to reverse-engineer. It retrieves live on essentially every query, always shows its citations, and weights recency and community content much more heavily than Google does.

If a Reddit thread from three weeks ago answers the question well, Perplexity will cite it. That makes it the fastest engine to move: a well-answered forum thread, a fresh comparison post, or a recently updated page can start showing up in citations within days rather than months. It's also the best diagnostic engine, because the citation list tells you exactly which sources are in play for your category. Even if Perplexity isn't your biggest traffic source (it reported 780M queries in May 2025, growing 20%+ month over month, which is large but far behind ChatGPT), it's the one to read when you want to know why an answer looks the way it does.

Google AI Overviews

AI Overviews grounds in Google's own index and leans heavily on pages that already rank. This is the engine where classic SEO transfers most directly, and also the one where the traffic math is most brutal.

Coverage is broad and skews toward exactly the queries B2B companies care about. BrightEdge found AI Overviews now appear on ~48% of tracked queries, up 58% year over year, and on 82% of B2B technology queries, 88% of healthcare and 83% of education. Semrush's more conservative panel puts it lower, at 15.69% of keywords in November 2025, having peaked at 24.61%, and the gap between those numbers is mostly a difference in which keywords each panel tracks. Either way, the direction is one-way.

And when an Overview appears, the click leaves. Ahrefs measured the #1 organic result losing 58% of its clicks, and Pew found only 8% of users click any result when an AI summary is present, versus 15% without, with just 1% clicking a link inside the summary.

Gemini and AI Mode

Gemini blends model knowledge with Google grounding, and Google AI Mode passed 1 billion monthly users by May 2026. AI Mode does more aggressive query fan-out than a standard AI Overview, which means it retrieves from a wider set of sub-queries and gives long-tail, use-case-specific content a better shot at being pulled in.

The reason this table exists is that it destroys the idea of a single "AI visibility score." Your ChatGPT mention rate and your AI Overviews mention rate are different numbers with different causes, and a tool that averages them into one figure is hiding the thing you need to see. Which is also why we think per-engine reporting is non-negotiable, and why it's worth checking what each tool actually covers before you buy: most of the category gates engines behind higher tiers.

What makes a passage citable?

From looking at a lot of cited passages, the ones that win share a small number of properties:

• The claim is checkable. A number, a date, a price, a named entity. "Spottlo tracks 25 prompts per brand across four engines for $39/month" is citable. "Spottlo offers comprehensive coverage" is not.
• It's self-contained. The passage makes sense lifted out of the page, with no antecedent pronouns pointing at a paragraph the model didn't retrieve.
• It answers the heading directly. Chunking follows heading boundaries. If the heading and the answer land in different chunks, neither chunk is useful.
• It's in a table or a list when the data is comparative. Tables survive chunking well and get cited disproportionately, because the structure is unambiguous.
• It's attributable. A named author with a real title, a publication date, and an original source for any statistic. Engines are increasingly conservative about grounding claims in anonymous content.

What to do next

1. Pick your ten highest-intent buyer questions and run them through Perplexity. Read the citation list. That list is the map of who the engines already trust in your category.
2. For each question you lose, log the cited domains. Sort by frequency. The top five are your outreach targets, and getting into them beats three months of publishing.
3. Rewrite your commercial pages answer-first, one checkable claim per heading, with tables where the data is comparative.
4. Verify AI crawler access for GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended. Retrieval is stage two, and blocking it means you never reach stage three.
5. Measure per engine, never averaged. Start with a free AI visibility report to see where the gaps are, then track the same prompt set weekly so you're reading a trend instead of a coin flip.