Six moves that actually get your brand named in an AI answer — and four popular ones that do nothing.
Your next customer might never type your name into Google. They’ll ask an assistant “what’s the best option for X” — and take the two or three names it hands back. If you’re not one of them, you never entered the conversation.
There is no page to optimise here, no meta description to tune, and no button that submits you to ChatGPT. What there is, is a chain: models name brands they find corroborated in sources they already trust. Work the chain and your AI visibility moves. Here are the six moves that do it, in the order they pay off — and, at the end, the four that everyone tries first and that do nothing at all.
Do this one first, because every other move on this list is wasted if it fails. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are four separate crawlers with four separate permissions, and none of them is Googlebot. A robots.txt written years ago for search engines, or a firewall rule that blocks unfamiliar bots, will quietly exclude the models you are trying to reach.
A blocked crawler is a mention rate that cannot move no matter what else you do — and it fails silently, because nothing anywhere tells you it happened. Check the file, and check it names each agent explicitly.
Keyword lists were built for a search box. Nobody asks an assistant “best CRM small business” — they ask “what CRM should I use for a six-person agency that hates admin”, and it is the answer to that sentence you are competing for.
So write the questions out in full, in your buyers’ words, and group them by intent: comparisons, “best for”, alternatives-to, and the awkward ones about price and drawbacks. Fifteen good questions beat a hundred keywords, because each one is a specific answer you can go and win.
This is the move most people skip, and it is the one that makes the rest tractable. An overall visibility score is an average, and an average is not a task. A single question you lose is.
Take the worst-performing question and treat it as the entire job for a fortnight. Ask each model, and note which ones name you and which don’t — they will disagree, often sharply, because they don’t all read the same internet. One model recommending you and another never mentioning you is not a contradiction to resolve; it is two different source problems, and the model that ignores you is telling you where the gap is.
If you can’t see which prompts you lose, you can’t win them back.
Read the answer you lose and list what it cites — not who it names, what it links. Models rarely cite your homepage. They cite Reddit and niche forums, review sites, directories, independent comparison posts and trade press. That citation list is the real target, and it will be different for a coffee roaster than for a payroll tool.
Then go and earn a place in those specific sources. In practice that is unglamorous and quite finite:
Your own site is not where the citation comes from, but it is what gets checked once a model has a reason to look. Make that easy: answer each question plainly, near the top, in a self-contained paragraph that makes sense lifted out of the page. State the specifics — prices, sizes, timings, who it’s not for — in text rather than in an image, behind a form, or built by a script the crawler never runs.
The test is simple: could someone paste one paragraph of this page into a chat and have it stand on its own? If yes, a model can quote it. If it only makes sense with the three paragraphs above it, it can’t.
Asking ChatGPT about yourself once tells you almost nothing. These models aren’t deterministic — ask the same question twice and you can get two different shortlists — and they disagree with each other by design. A single check is a coin flip you will read as a verdict.
Measure the same questions, across the same models, on a fixed schedule, and read the trend rather than any one run. It also sets expectations honestly: sources have to be published, then crawled, then absorbed, so the work you do this month usually shows up in the number weeks later. That lag is normal, and it’s the reason to track rather than to keep checking.
Worth naming, because they absorb most of the effort we see spent here:
The uncomfortable summary is that there isn’t a growth hack here yet. Getting recommended by AI looks a lot like being genuinely well-regarded in public, which is slow — but it also means the brands winning today aren’t louder than you. They’re just the ones watching the number while everyone else guesses.
Get cited in the sources it already reads. Models name brands they find corroborated in places they trust — Reddit threads, review sites, directories, independent comparison pages — so the work is earning a place in those, not editing your own homepage. Check the AI crawlers can reach your site first, because nothing else moves until they can.
Not in the organic answer. Buying mentions in bulk on low-quality sites is counterproductive, because those are exactly the sources models weight down. Paid placement inside AI products exists as advertising, and it is labelled separately from the recommendation itself.
Weeks, not days. A source has to be published, then crawled, then absorbed into what the model will repeat — so work done this month typically shows up in the number some weeks later. That lag is why visibility is worth tracking on a schedule rather than spot-checking.
No. Structured data helps a model parse a page it already reached; it does not make the page get cited, and there is no tag that asks to be recommended. It is worth doing for other reasons — just not this one.
Because they do not read the same internet. Some models search the live web and cite as they answer; others reply from training data. A model that ignores you is usually pointing at a specific sourcing gap rather than disagreeing with the others.
Mention Tracker puts fifteen of your buyers’ questions to ChatGPT, Claude, Gemini, Perplexity and Grok every week, and lists the ones you lose worst first — with the sources the winning answers cited.