How AI Search Engines Decide What to Recommend (And How to Get on the List)

When someone asks ChatGPT "what's the best CRM for a small law firm," the answer names two or three products. Somebody decided that. Not a person — but a process, and the process is knowable enough to work with.

Here's what determines whether you're on that list, in the order it gets decided, with what you can actually do about each.

First, drop the ranking metaphor

There is no ranked list of results with you sitting at position 14. The system assembles a fresh set of candidate sources for each question, narrows it, then writes an answer naming a handful. Three consequences worth internalizing:

There's no position to hold. You can appear in nine phrasings of a question and vanish in the tenth. One check is not a measurement
Being "first" isn't the goal. The goal is being named at all. Effort spent on relative position is spent on something that doesn't exist
Different failures need different fixes. Never retrieved, retrieved but never picked, and picked but never named are three separate problems with three separate solutions

Factor 1: Can a machine read your pages?

Binary, and still a surprisingly common cause of total invisibility. If your content is rendered by JavaScript after load, sits behind a login, or is blocked by an interstitial, it isn't in the candidate pool and nothing else on this page applies.

Check it: open your key page with JavaScript disabled in your browser. If the text isn't there, that's your entire problem right now.

Fix it: server-render or pre-render the content that matters. This is an engineering ticket, not a marketing one, and it should jump the queue.

Factor 2: Do you use your customers' words?

Candidate sources get pulled by meaning, not exact keywords — but it's still matching. Pages written in internal vocabulary retrieve poorly against questions written in customer vocabulary.

This is the quietest failure mode we see: the page is crawlable, well-written, professionally designed, and phrased in language no customer would ever type. "Unified revenue orchestration platform" does not retrieve against "software to stop losing track of invoices."

Check it: pull twenty real questions from support tickets and sales calls. Search your own site for the nouns and verbs in them. If they're absent, you have a vocabulary gap.

Fix it: write the customer's phrasing into your headings, then bridge to your terminology. Both can coexist — one gets you retrieved, the other keeps your positioning.

Factor 3: Can the answer be lifted out cleanly?

This is the big one, and the biggest departure from SEO habits.

Once there's a pool of candidates, it gets narrowed to what fits in the answer. What survives is content where the answer sits in a self-contained block — a passage that states one specific thing and still makes sense pulled out of context.

A well-known brand with the answer spread across six paragraphs loses to a smaller competitor with one tight paragraph. This is exactly how big companies get beaten by startups in AI answers while still outranking them in Google.

Check it: take any 80-word window from your page. Paste it into a blank document. Does it still assert something specific and complete? Most marketing copy fails this everywhere on the page.

Fix it: one question per heading, answer in the first sentence beneath it. Claim before support, always.

Factor 4: Does anyone else say it about you?

Systems weight third-party description heavily — and they weight it differently from links. A sentence on someone else's credible page saying "X is a provider of Y for Z" is text the engine can reuse directly. A backlink with no describing sentence isn't.

This is why the same claim, made only by you, loses to a claim made by you and three other sources. And it's why you can't buy your way here in any way that lasts.

Check it: search for your company name and read what other sites say you do — not whether they link to you. Count the pages containing a describing sentence.

Fix it slowly: analyst coverage, trade press that describes rather than announces, partner documentation, community answers, integration listings. Not press releases — announcements describe events, not identity, which is why a busy newsroom can generate almost no corroboration.

Factor 5: Is it safe to name you?

The last filter is narrower than the others, and it's the one almost nobody works on: an answer uses more sources than it names. Getting named is a separate decision from getting used.

What makes a source safe to name: an identifiable author or organization, a visible date, a stated basis for the claim, and specifics rather than generalities. Anonymous undated prose can be useful to an engine and still never worth pointing a user toward.

Check it: when you do appear in answers, note whether you're credited or just described. If your framing and numbers show up with no credit, you're being used and not cited — a good position to be in, and a fixable one.

Fix it: put a named author, a date, and a one-line methodology on anything making a factual claim. See the full citation checklist.

A useful way to hold all five

GEO: The New Enterprise Gateway in the AI Era (Huang Shuang, Li Jing, Wu Hao, Luo Wei, Yuntao Zou) frames this as three stages: being seen → being understood → being chosen. Factors 1 and 2 are being seen. Factors 3 and 4 are being understood. Factor 5 is being chosen.

The reason the framing is worth borrowing: most brands measure only the first stage and assume the rest follow. They don't. Being retrieved reliably, selected occasionally, and named almost never is a completely normal state — and it looks like success if all you track is whether you show up somewhere.

Common questions

Which factor should I fix first?

In order. Factor 1 is binary and cheap — do it this week. Factor 2 is a week of editing. Factor 3 is the highest-leverage content work. Factor 4 takes months. Factor 5 is a small change with outsized effect once 1–3 are done.

Do different AI platforms behave differently?

Yes, enough that a single-platform check misleads. The same brand can be well represented on Perplexity and nearly absent from ChatGPT. Test at least four — our 60-day tracking found platform divergence larger than most week-to-week changes.

Is there anything I can pay for that works?

Not for factor 4 in any durable way. Paid placement that produces a genuine describing sentence on a credible page has some effect; link buying has essentially none, and reads as a manipulation signal.

How long before I see change?

Roughly twelve weeks before a real change is distinguishable from noise, assuming you fixed the right factor. Faster than that is usually you measuring variance.

Does publishing more help?

Not by itself. Volume affects how often you're crawled, not whether you're chosen. Six substantive pieces beat sixty thin ones.