AI search visibility

Why rankings no longer tell you the whole story

A growing share of commercial questions never produce a click. The buyer asks an assistant, reads a synthesised answer, and either remembers three brand names or does not. If you are not one of those three, you were never in the consideration set, and no ranking report will tell you that happened.

This is why impressions can stay flat while a category quietly reallocates demand. The traffic you lose to answer engines does not show up as a ranking drop — it shows up as fewer sessions at the same position.

The three metrics worth tracking

Most tools report one number. Three are needed to act:

• Mention rate — the share of probe questions where your brand name appears in the answer at all. This measures whether the model knows you exist.
• Citation rate — the share where your own URL is listed as a source. This measures whether your pages are retrievable and quotable, not just whether the model has memorised your name.
• Leakage — which competitor took the slot when you were omitted, and for which question. This is the single most actionable number, because it names the page you need to write.

Building a probe set that means something

Write 20 to 40 questions the way a buyer would type them, not the way a marketer would. Include category questions ('best tool for X'), comparison questions ('X vs Y'), problem questions ('how do I stop Z happening'), and one or two direct brand questions as a control.

Run the same set on the same schedule — weekly is enough — against every model that matters to you. Never change the questions mid-experiment, or you lose the trend. Record the full answer text, not just yes/no, so you can see how you were described when you were mentioned.

What actually moves the numbers

In order of observed impact: fixing entity confusion (the model mixing you up with a similarly named company), making pages crawlable as plain HTML, adding answer-first blocks with checkable facts, publishing genuinely comparative content that names competitors honestly, and getting mentioned on third-party pages the models already retrieve.

What does not reliably move them: keyword density, publishing volume without structure, and adding schema to a page with nothing quotable in it.

What we measured on real domains

Everything below comes from ThoughtForge runs against live domains, not from a study we read. Each run scores three things out of 100: identity (can an engine say who you are without contradicting itself), durability (does your brand survive a multi-turn conversation), and retrieval (can your pages be chunked into something quotable).

Across every domain we have scanned, one number is consistently the worst, and it is not the one people optimise for. Durability sits at or near zero almost everywhere. Brands that get named in turn one disappear by turn three, because nothing on the site gives the model a reason to keep holding onto them.

• thoughtforge.app, July: identity 16, durability 16, retrieval 49, overall 27.
• thoughtforge.app, September, after entity and schema work: identity 52, durability 0, retrieval 48, overall 33.
• A services domain (megap•••••••••••.com.au): identity 28, durability 0, retrieval 47, overall 25.
• An industry association (propertyasso•••••••.com.au): identity 4, durability 0, retrieval 50, overall 18.
• A consultancy (obse•••••••••.com): identity 40 in July, 52 in September; overall 30 to 35.

What actually moved the numbers

Identity is the score that responds fastest. On thoughtforge.app it went from 16 to 52 in about six weeks, and the work was unglamorous: one Organization schema block that agreed with itself, one description of the product used in the same words on the site, in the schema and on every third-party listing, and crawler access opened to GPTBot, ClaudeBot and PerplexityBot.

Retrieval barely moved — 49 to 48 — because retrieval is about page structure, not metadata, and we had not restructured the pages yet. That is the honest shape of this work: metadata fixes buy you identity, and only rewriting the pages buys you retrieval.

Durability moved backwards, from 16 to 0. Multi-turn persistence is the hardest of the three and the one almost nobody is working on. If your brand only needs to be named once, ignore it. If you are in a considered purchase where the buyer asks the assistant four or five follow-ups, it is the whole game.

How to run your first visibility measurement

Pick five questions a buyer would actually type — not your brand name, the problem they have. Ask each one in ChatGPT, Claude, Perplexity and Gemini, and record whether you are named, which competitors are named instead, and which sources get cited. That manual pass takes about forty minutes and is more honest than most dashboards.

Then push each conversation three turns further. Ask for a comparison, then a recommendation, then a justification. This is where most brands vanish, and it is invisible if you only ever ask one question.

Reading the result without fooling yourself

Two failure modes look identical from a single answer. In the first, the engine cannot retrieve your pages at all. In the second, it retrieves them and chooses not to quote them. The remedies are opposite — access and structure versus substance and specificity — so separate them before you act.

The tell is citations. If your domain never appears as a source anywhere in the conversation, you have a retrieval problem. If it appears as a source but the summary describes you vaguely or wrongly, you have an identity problem.

Frequently asked questions

How is AI search visibility different from share of voice?

Share of voice measures mentions across media. AI search visibility measures mentions and citations inside generated answers to specific buying questions, which maps far more directly to demand.

How often should I probe?

Weekly is enough for most categories. Daily adds noise without adding signal, because model answers vary run to run.

Do AI Overviews count?

Yes, and they are worth tracking separately. Google AI Overviews draw on the classic index heavily, so they respond faster to conventional SEO fixes than ChatGPT does.

Related guides

• How to rank in ChatGPT — ChatGPT does not rank pages, it picks sources to cite. Here is what makes a page citable, how to measure it, and the fixes that change the answer.
• AEO vs SEO vs GEO — SEO wins rankings, AEO wins the answer, GEO wins the citation. What each term actually means, how they overlap, and how to measure all three.
• Why AI doesn't mention your brand — Six concrete reasons ChatGPT and Perplexity skip your brand, how to tell which one applies to you, and the fix for each.