The two numbers that matter in AI search: trust and visibility.

When someone asks ChatGPT, Gemini, Perplexity, Claude or Grok which product to use, the engine does not consult a league table. It decides whether it needs to search, rewrites the question into queries of its own, reads a handful of pages, decides which of them to believe, and writes one answer with a few names in it. A brand can drop out at any of those steps, and the fix is different at each one.

That pipeline is the whole subject. Everything else, the schema files, the special text files for crawlers, the domain-rating charts, is either a small part of one step or nothing at all. Google says outright that it does not read the llms.txt file, and that no special markup is needed for its AI features. What moves the outcome is what the engine can retrieve, what that material says, and who else says it.

Measure two things, and keep them apart

Buyers ask engines two very different kinds of question. They ask about brands they already know, and they ask about categories they are trying to choose in. These need separate numbers, because a brand can be described beautifully when named and be entirely absent when the buyer does not know to name it.

Trust is what the engine says when a buyer checks you by name. Not sentiment. An answer can sound warm and still say "there is limited independent evidence about its security", which is a trust failure with a positive tone. The useful measure is whether the answer is accurate, current, evidenced by sources that are not you, fair about risks, and confident enough to recommend.

Visibility is whether you enter the consideration set when a buyer asks a category question without your name in it. The word that matters is qualified. A mention in the wrong category, or as the example of what not to buy, counts for nothing. Underneath the headline sit mention rate, recommendation rate, first-recommendation share and share of voice against the rivals actually being named.

Citations, the links under an answer, are a diagnostic and not a goal. Large 2026 datasets show engines citing a site without naming the brand, and naming a brand without citing its site, so often that the two barely correlate. Track both. Never add them together.

The engines are not one system

The single biggest mistake in the field is a plan written for "AI". Each engine has its own way of reading the web and its own diet of sources, so the same brand needs a slightly different footprint for each.

Two consequences follow. First, report per engine; a blended share of voice hides a win on Perplexity behind a loss on ChatGPT. Second, off-site presence matters more than on-site polish. Across 75,000 brands, mentions on YouTube and around the web correlated with AI visibility far more strongly than domain rating or the size of the site. The engines trust what other people say about you.

Name the failure before you fix anything

Because the pipeline has steps, an absent brand has a reason, and the reason picks the work. We use twelve failure classes. The ones that come up most:

• F1Prior lag. The engine's default answer is stale but its search-on answer is right. The web is fine; the model has not caught up. Feed the ecosystem, and never promise a date.
• F2Eligibility. The engine's own retrieval bot cannot read your pages. Every engine has a different bot and a different robots.txt contract; test what each one actually receives.
• F4Selection. You have a relevant page, but rivals' sources keep winning the retrieval. The fix is on the sources the engine already reads, not on your page.
• F6Corroboration. Only you say it. Owned pages state a claim; independent sources make it believable. This is the trust mechanic every engine shares.
• F8Narrative. You are included and described wrongly, usually as the product you used to be. Fix it at the source with a dated page that says what you are today and what you no longer are.

The order of work follows from this. Get the facts about the brand straight, with evidence. Make sure each engine can retrieve them. Make every representation of the brand agree, on your site, in the category's databases, on the profiles engines resolve identity from. Find the ten or twenty pages that already shape the category's answers. Then, and only then, write the smallest set of resources that covers the buyer's decision, and earn presence on the sources that were shaping the answers without you.

Measure like a scientist, not a dashboard

Answers are stochastic. The same prompt, run five times, does not give the same answer five times. A single Monday reading that says 100% and then 0% has told you almost nothing. Freeze a benchmark set of prompts, run each several times per engine, report the variance with the number, and count only movement that holds beyond it. Record whether the engine searched or answered from memory, because those are different results. When you change the prompt set, start a new series rather than pretending the old one continues.

Do that, and the two numbers become something you can act on: a trust reading that tells you which claim lacks proof, and a visibility reading that tells you which sources to earn. That is the whole method. The rest is doing the work.