You launched. Someone on the team asks ChatGPT for the best tool in your category. It names 3 competitors. Your product isn’t there.
The next move feels obvious. Fix the website: add an FAQ, add schema, add an llms.txt file. Or reply to one of the agencies that offered a GEO package with a guaranteed place in AI answers.
The reasoning makes sense. If AI can’t find you, make yourself easier to find. Sometimes that is exactly the problem: the site blocks AI crawlers, or the text loads only through JavaScript and the crawler sees an empty page.
More often the site is fine, and the answer still names someone else. The answer was built from other pages.
When someone asks ChatGPT, Perplexity or Google’s AI for a product, the assistant usually searches the web first. One question turns into several smaller searches. It reads some of the pages that come back and writes the answer from them. The brands on those pages are the brands it names.
Follow one question.
Someone asks an assistant for a product.
The assistant splits the question into several searches.
Each search brings back pages. It reads some of them.
It names the brands those pages mention.
3 brands named. All 3 appear on at least 2 of the pages it read.
Your site is one of the pages it might read. Most of the others belong to someone else.
There is a second source: what the model remembers from training. That memory is months or years old and changes when a new model comes out. For a new product, the live search is where you can move something now.
We asked ChatGPT, Copilot and Google’s AI the questions customers ask about 65 local firms in Italy: window makers, wedding venues, pool builders. Then we looked at which pages each answer was built from.
For 2 firms in 3, a directory was the most-cited source of all.
The sites still mattered. Firms whose site never appeared in the answers were named about 3 times less often.
Your site gets you read. Other people’s pages get you named.
Partly, yes. Being written about has always mattered. 2 things are new.
Start with the question. Is there a question your buyer asks where you could be the best answer? “Invoice matching for agencies with fewer than 20 clients” has a handful of good pages on the web. “Software for documents” has thousands.
The question that changes the next move: which narrow question do we want to be named for, and which pages answer it today?
This is the most common picture. Ask “what is [your product]” and the assistant finds your site. Ask what your buyer asks, without your name, and you’re missing. The first result flatters you. The second is where buyers choose.
The question that changes the next move: which pages name our competitors for our buyer’s questions?
Wrong category, old price, a buyer you don’t serve. The assistant repeats what the pages say. If a directory or a reference entry files you under the wrong category, you can drop out of answers about the right one.
The question that changes the next move: where does the wrong description come from?
On “the best CRM” they will keep winning: they fill every review site and comparison list. That race is closed to a new product for now.
The question that changes the next move: which narrower question sits inside the broad one, where our buyer is more specific than the big brands’ pages?
Ask the same assistant the same question twice and you get 2 different answers. SparkToro, a market research company, found less than a 1 in 100 chance of getting the same list of brands across 100 runs. A screenshot of one answer proves nothing. “We’re #1 in ChatGPT” means nothing either.
So measure a share. Take questions your buyer asks without your name, run each one several times, and count how many answers name you.
Track 3 numbers.
Then check them against the number that pays: how many new leads say they found you through an assistant.
4 weeks of work. Don’t expect the share to change in the same month: a new page counts only once assistants find it, and the model’s memory changes only with new models.
5 checks. A developer does them in an afternoon.
robots.txt. At yoursite.com/robots.txt, allow OAI-SearchBot (ChatGPT search), PerplexityBot and Googlebot. GPTBot is a separate switch: it decides whether your pages are used for training.
Bing. ChatGPT search works with other search providers, so allow Bingbot and submit your sitemap in Bing Webmaster Tools.
Your CDN. Since July 2026 Cloudflare splits AI bots into Search, Agent and Training, each with its own switch. Keep Search allowed, and Agent too: it covers ChatGPT opening your page when a user asks. Blocking Training also blocks crawlers that do both jobs, such as Googlebot and Bingbot.
Text in the page. Open “View page source” and search for your main sentence. Most AI crawlers don’t run JavaScript: if the text appears only after the page loads, they see an empty page.
No special files. Google lists no requirement for llms.txt, AI text files or extra markup to appear in its AI answers.
Answers stay in your browser until you choose to send them.
If the map shows the question and the pages, start the 4 weeks. If you couldn’t write 5 questions your buyer asks, the open question sits before GEO: whom the product is for.
A Strategic Reframe Session is 60 minutes with a Beyonds mentor on your whole project. Being named in AI answers is one channel. Which channel comes first depends on where the project stands:
You leave with 3 documents written for your project: where it stands, the risks ahead and what to do next. If you already know your buyer and your channels, skip the session and start week 1.
Your AI Visibility Map travels with the request, so the mentor sees your buyer and your questions before the call.
If the open question is whether the market will move before a full build, Market Pull Route has its own free selection session.