How AI Search Engines (Gemini, ChatGPT, Perplexity) Decide Which Brands to Recommend
AI search engines do not rank brands. They assemble an answer, and brands get named as a side effect of that assembly. Understanding the difference is the whole game: you are not competing for a position, you are competing to be the obvious thing to say when a model writes a sentence.
When someone asks Gemini, ChatGPT or Perplexity "what's the best tool for X", the engine runs a short pipeline — interpret the question, decide whether to search, retrieve candidate sources, select which of those to trust, then synthesise a fluent answer. Brands survive that pipeline when they are unambiguous, corroborated by several independent sources, current, and easy to retrieve.
This guide walks through that pipeline step by step, explains where the brand list actually comes from, shows how the major engines differ, and separates the signals you can influence from the ones you cannot.
Key takeaways
- A recommendation is produced by a pipeline, not a ranking algorithm. Different failures happen at different stages.
- Brand names come from two places: pretrained knowledge (slow, hard to change) and live retrieval (fast, very changeable). Diagnose which one is failing you.
- Source selection and brand selection are different problems. Being retrieved is not the same as being recommended.
- Engines converge on the same few brands because they reward consensus — which is also the mechanism a challenger can exploit.
- The most controllable signals are entity clarity, retrievability and recency. The least controllable is accumulated reputation.
The five steps between a prompt and a recommendation
Step 1 — Interpretation
The engine parses what you actually want. A prompt like "I run ops at a 30-person agency and need something to track brand mentions in AI answers — what should I look at?" is decomposed into a use case, a company size, a category and an implicit budget range.
Why it matters for you: the engine is matching against a category, not a keyword. If your positioning is vague — if you describe yourself as a "growth platform" rather than an AI visibility tracker — you may not be considered a candidate at all, no matter how good your content is.
Step 2 — The search decision
The engine decides whether to answer from memory or go and look. Perplexity almost always searches. ChatGPT searches when the question implies recency, specifics or products. Gemini leans on Google's index heavily.
Why it matters: this single branch decides which of your assets is in play. Answered from memory, your content this quarter is irrelevant — what matters is what the open web said about you over years. Answered from search, a page you published last week can be cited today.
Step 3 — Retrieval
The engine pulls a candidate set of documents. This is a fairly conventional search operation: relevance, authority, freshness.
Why it matters: if AI crawlers cannot render your pages, you are absent from the candidate pool before any judgement about quality happens. Most AI crawlers execute JavaScript far less reliably than Googlebot. Fetch your own page with curl — if the article text is not in the raw HTML response, it is not in the model's context either.
Step 4 — Source selection
From the candidate set, the model chooses which sources to actually lean on. It favours documents that are specific, well-structured, internally consistent and mutually corroborating. A page that makes a concrete claim also made by three other independent sources will be trusted over a page making a claim nothing else supports.
Why it matters: this is where most "we have great content" brands lose. One excellent page on your own domain is a single source, and a self-interested one.
Step 5 — Synthesis
The model writes an answer. Here it decides how many brands to name (usually three to five), in what order, and with what hedging. Being named first with a clear recommendation is a very different outcome from being appended to a list with "other options include".
Why it matters: your goal is not merely presence. It is being the brand the model feels confident enough to name first.
Where the brand list actually comes from
Two sources, with completely different optimisation logic. Confusing them is the most common reason GEO effort is wasted.
| Pretrained knowledge | Live retrieval | |
|---|---|---|
| What it is | What the model absorbed during training | What it finds by searching mid-answer |
| Update speed | Months to years | Minutes to days |
| What influences it | Accumulated coverage across the open web | Crawlable, relevant, current pages |
| Can you edit it? | Not directly | Yes, quickly |
| Typical failure | "The model has never heard of us" | "The model can't read our site" |
| Fix | Third-party corroboration, sustained coverage | Technical access, answer-first content |
The diagnostic that matters: ask an engine your category question twice — once with web search disabled or in a mode that answers from memory, once with search active. If you appear only in the second, you have a memory problem and need years of corroboration compressed into a deliberate PR and community programme. If you appear in neither, start with retrievability, because that is fixable this month.
What makes a source get selected
At the retrieval and source-selection stages, engines consistently favour documents with these properties:
| Property | What it looks like in practice |
|---|---|
| Self-contained sections | A passage makes sense when lifted out of the page, with no "as we mentioned above" |
| Specific claims | Numbers, definitions, named mechanisms — not adjectives |
| Clean structure | Descriptive headings, short paragraphs, real tables |
| Consistency | The page does not contradict itself or your other pages |
| Freshness signals | Visible dates, dateModified, current facts |
| Third-party framing | Comparison and roundup content is retrieved heavily because it is explicitly evaluative |
Retrieval works on chunks, not pages. The unit that gets pulled into the model's context is a section, sometimes a few hundred words. Write accordingly — the tactical detail is in how to optimize your content for Gemini and ChatGPT recommendations.
What makes a brand get named
Source selection gets your page into the context window. Brand selection is a separate hurdle, and it depends on properties of your brand rather than your page:
- Entity clarity — the model can state precisely what you are, and every source agrees.
- Corroboration — several independent sources describe you the same way.
- Category fit — you appear in the third-party lists that define your category.
- Recency — the information about you is current; stale facts actively suppress mentions.
- Sentiment — you are described favourably, not merely acknowledged.
- Risk — the model is not exposed by recommending you. Ambiguity reads as risk.
These are unpacked, ranked by how much you can control them, in 7 key factors that influence your brand's visibility in LLMs.
The uncomfortable implication: most of what determines whether you get named was written by someone else. You can control your site completely and still lose, because the model is weighing a consensus you did not author.
How the major engines differ
| Engine | Search behaviour | Grounding source | What it rewards | Best way to test it |
|---|---|---|---|---|
| ChatGPT | Mixed — memory, browsing, and stored user memory | Its own index plus live search | Broad, consistent presence over years plus current pages | Compare answers with and without browsing |
| Gemini / AI Overviews | Search-first | Google's index and Knowledge Graph | Classic SEO fundamentals, structured data, strong entity signals | Check whether you rank for the underlying query |
| Perplexity | Always searches, always cites | Live web retrieval | Directly quotable, well-structured, current pages | Read the citation list — it shows you the retrieval pool |
| Claude | Conservative, searches when needed | Live search, favours primary sources | Documentation, primary sources, credible third-party analysis | Ask for sources explicitly |
| Copilot | Search-first | Bing index plus enterprise context | Business sources, review platforms, technical docs | Test with business-framed prompts |
Start with Perplexity. It shows its sources, which makes it the only engine that will tell you why it named your competitor. That citation list is a fair approximation of the retrieval pool everyone else is drawing from too.
Why the same three or four brands keep appearing
Ask any engine for recommendations in a mature category and you will usually get the same small set. This is not a bug, and it is not a paid placement. It falls out of how the pipeline works.
Consensus is a proxy for safety. A model producing a recommendation is managing risk. The brand that twenty independent sources describe in consistent terms is a low-risk thing to say. A brand described by two sources, inconsistently, is a risk — so it gets omitted, hedged, or replaced by a safer name.
Retrieval pools are shallow. For a given question, the engine may pull a handful of documents. If the same five roundup articles dominate your category's search results, those five articles effectively define the candidate brand list.
Training data has gravity. Brands that were widely written about during the training window carry forward, even when the market has moved on.
Read as a challenger, this is more encouraging than it sounds. The concentration is not driven by an unassailable moat — it is driven by corroborated description, which is addressable. Get into the roundups that dominate your category's retrieval pool, get described consistently across review platforms and communities, and you are altering the exact input the pipeline runs on. That is a far shorter path than accumulating a decade of domain authority.
What you can and cannot control
| Signal | Control | What to actually do |
|---|---|---|
| Retrievability | Full | Allow AI crawlers; serve real HTML; verify with curl |
| Entity clarity | Full | One consistent description everywhere; complete Organization schema with sameAs; a real About page |
| Factual accuracy | Full | Keep pricing, features and integrations current — stale facts get repeated to prospects |
| Content structure | Full | Answer-first sections, specific claims, tables |
| Category fit | High | Get into the third-party roundups that define your category |
| Review presence | High | G2, Capterra, Trustpilot profiles that are complete and current |
| Corroboration | Medium | Digital PR, community participation, original data others cite |
| Sentiment | Medium | Address the substance behind negative coverage; see sentiment and credibility in AI search visibility |
| Training data | Low | Only via sustained coverage over long periods |
| Model architecture | None | Ignore it; optimise the inputs |
The honest summary: you control roughly the first half of this table completely, and most brands have not finished it. Entity cleanup and crawler access are unglamorous and frequently skipped, and they gate everything below them.
What this means if you are the challenger
The pipeline description above can read as discouraging — a system that rewards incumbency, weighted towards brands with a decade of accumulated coverage. That reading is half right, and the other half is where the opportunity is.
Incumbency helps at the memory stage. It helps much less at the retrieval stage. When an engine searches mid-answer, it is evaluating documents that exist right now, on their merits: specificity, structure, recency, corroboration. A three-year-old company with a genuinely better comparison page can win that competition against a fifteen-year-old company whose equivalent page was last updated in 2022.
Four moves that work disproportionately well for challengers:
Compete on narrowness. Broad category prompts ("best CRM") favour incumbents heavily. Specific ones ("CRM for a two-person recruitment agency that needs LinkedIn sync") have shallow retrieval pools, because almost nobody has written seriously about them. The buyer asking the specific question is also worth more to you.
Own a definition. Categories accumulate terminology that nobody has defined cleanly. Write the clearest definition of a term your market uses loosely, use it consistently, and it becomes the version that circulates — with attribution.
Publish the comparison your competitors will not. Incumbents rarely publish honest head-to-head comparisons, because they have more to lose. Doing it credibly — including where a competitor genuinely wins — makes you a source for comparison queries rather than a subject of them.
Report original numbers. Even small-scale testing produces data nobody else has. A dataset of ten prompts across three engines, honestly described, is more citable than another restatement of the category consensus.
None of this changes the model's default memory quickly. All of it changes the retrieval pool this quarter.
Three common misreadings
"There must be a ranking algorithm we can reverse-engineer." There is a pipeline with several distinct stages, and your failure is usually located at one of them. Treating it as a single ranking system leads teams to optimise the wrong stage — polishing content when the actual problem is that PerplexityBot receives an empty page.
"If we publish enough, we will get picked up." Volume does not address the two hardest inputs: corroboration and entity clarity. Twenty thin articles move nothing that four solid ones plus a cleaned-up entity footprint would not have moved further.
"The engines are biased towards brands that pay them." The organic recommendation is generated from retrieved and learned information. What looks like paid preference is usually consensus concentration — the same well-corroborated brands dominating a shallow retrieval pool. That distinction matters, because consensus is addressable and payment would not be.
How to test the selection logic yourself
An afternoon's work that will teach you more than any framework:
- Write ten prompts a serious buyer would type — full sentences with context, not keywords, and without your brand name in them.
- Run each on Perplexity, ChatGPT and Gemini, three times each, from a clean session.
- Record which brands are named, in what order, and with what framing.
- On Perplexity, record which URLs are cited. This is your category's retrieval pool.
- Look for the pattern: are the same domains cited repeatedly? Are you in them?
Two outcomes. Either you appear, in which case you have a position to defend and should measure it properly — see what is Share of Model and how to measure it. Or you do not, and you have just identified the specific third-party sources you need to be present in.
One caution: run each prompt more than once. These systems are probabilistic. Ask the same question three times and you may get three different brand sets. A single favourable answer is a coincidence, not a benchmark.
Your first 30 days
The pipeline gives you a natural work order: fix the earliest stage where you are failing, because a failure there makes everything downstream irrelevant.
| Week | Stage you are fixing | What to do |
|---|---|---|
| 1 | Retrieval | curl your key pages and confirm the text is in the raw HTML. Check robots.txt for AI user agents. Fix anything blocking access. |
| 1–2 | Interpretation | Write one plain sentence describing what you do and which category you are in. Deploy it verbatim on your site, schema, social bios and review profiles. |
| 2–3 | Source selection | Rewrite your five most commercially important pages answer-first, with specific claims and at least one table each. |
| 3–4 | Source selection | Correct every stale fact you can find — your pricing page, your integrations list, your third-party listings. |
| 4 | Measurement | Run ten prompts across three engines, three times each, from a clean session. Record brands named and sources cited. This is your baseline. |
Nothing on this list requires another team's cooperation or a budget approval, and together they address the four stages you control outright. The harder work — corroboration, category fit, sentiment — belongs to the quarter after this one, and it produces much better returns once the foundation underneath it is sound.
Frequently asked questions
Can you pay to be recommended by an AI search engine?
Not in the organic answer, no. Advertising formats are appearing around AI answers, but the recommendation itself is generated from retrieved and learned information. Attempts to manipulate it directly — hidden text, instruction-like content aimed at crawlers, manufactured review campaigns — are detectable, increasingly filtered, and risk the credibility that the whole mechanism depends on.
Why does the answer change every time I ask?
These systems are probabilistic and often retrieve live, so outputs vary between runs. This is why measurement requires repeated sampling with a fixed prompt set, and why a single screenshot proves nothing.
How long does it take to change what an engine says about us?
Retrieval-layer changes can appear within days of publishing or fixing a page. Changing the model's default answer — what it says without searching — depends on accumulated coverage and takes considerably longer. Plan in quarters.
The engine is saying something wrong about us. What do we do?
Correct it at the source rather than complaining about the output. Identify which pages the answer is grounded in (Perplexity will show you), update your own canonical pages, get outdated third-party listings corrected, publish a clear current statement of the fact, then re-test in a few weeks.
Is this different from SEO?
The underlying infrastructure overlaps heavily — generative engines are grounded in search indexes — but the unit of competition shifts from the page to the brand, and the measurement layer is entirely new. See GEO vs SEO: how search is changing in the AI era for the full comparison, and our complete guide to Generative Engine Optimization for the discipline as a whole.
Find out what the engines say about you — free
Everything above is testable, and you should test it before deciding what to fix. The manual version takes an afternoon and covers ten prompts on three engines. The problem is that ten prompts is not a representative sample, and one run is not a measurement.
Geomyze does it properly: it builds intent-based prompts for your brand across purchase, discovery and comparison intents, runs them repeatedly against the major AI engines, and shows you which brands appear in your place and why.
Run your first AI visibility report free → — one full report, no card required. You will see your Share of Model, the competitors named instead of you, a breakdown by intent, and a prioritised list of what to fix first.
