Ask ChatGPT about your company. Then fix what it got wrong.
Buyers now ask ChatGPT, Perplexity and Google’s AI mode who does what you do. The answer comes from your site, in ways you can check and change. Twenty questions, what the answers reveal, and the fixes that moved them.
Open ChatGPT and ask it who fixes slow WordPress sites for B2B companies in your country. Then ask it what your company does and what it costs. Then ask Perplexity the same two things. The answers come from your website, filtered through whatever the crawler could read and whatever your pages said plainly enough to be quoted. On most sites I test, at least one of the twenty questions I ask comes back wrong, and the wrong answer is usually the site’s fault: a blocked crawler, a service page that never says what the service is, a price that lives in an image.
- Ask twenty real buyer questions in three assistants and write the answers down. That is the baseline.
- Check what the crawlers can read: robots.txt, the HTML before JavaScript, the challenge page they might be hitting.
- Say what you do in words a machine can quote: service, for whom, price, where. Schema repeats it in a form they parse.
- Measure again in six weeks, side by side. AI answers move, so the comparison is the result, not one screenshot.
Where the answer comes from
Three places, and you control two of them. The assistant’s training data, which is old and not yours to change. Live retrieval, where ChatGPT search, Perplexity and Google AI mode fetch pages at question time, read them, and quote the clearest one. And third-party pages about you: directories, reviews, a partner’s case study. The live retrieval is where the work is, because it is where your own pages are read, and where a crawler that can’t get in means you don’t exist.
When an assistant searches the web for a question, reads a handful of pages and composes the answer from them, with citations. It is how ChatGPT search, Perplexity and Google’s AI Overviews work. Your page is in the answer if the crawler could read it and if it said something quotable.
The twenty questions
I use the questions a buyer would actually type, not the ones a marketer would. Five about the category (“who does X for companies like us”), five about you by name (“what does Company do”, “how much does Company charge”), five comparisons (“Company vs competitor”), five problems (“our site is slow on mobile, who can fix it”). Each one in ChatGPT, Perplexity and Google AI mode. Sixty answers, in a sheet, with the sources each one cited.
The sheet tells you three things at once: whether you appear, whether what it says is true, and which page it read to say it. The third column is where the fixes come from. If every correct answer cites the same page and every wrong one cites a directory from 2019, you know exactly what to write.
20Buyer questions, asked in three assistants
60Answers in the baseline sheet, each with its sources
6 wkBefore the second run; sooner and the answers haven’t moved
What the wrong answers usually reveal
- The crawler never got inA security plugin or a CDN rule blocks anything that isn’t a browser, or robots.txt blocks GPTBot and takes the search bots with it. The assistant falls back to third-party pages, which are old. The scan on this site tests this with the real user agents.
- The page is empty until JavaScript runsA JavaScript-rendered site sends a shell and fills it in the browser. Google renders it, most AI crawlers do not. They read the shell: a logo and a cookie banner.
- The service page never says what the service isA headline that reads “Unlock your potential” and three paragraphs about values. A human infers; a model quotes. If the page does not contain the sentence “we do X for Y, from $Z”, nothing can quote it.
- The price is in an image, or nowherePricing tables rendered as screenshots, or hidden behind “get a quote”. The assistant answers the price question with a competitor who published one.
- Nothing about you is structuredNo Organization, Service or FAQ schema. The pages read fine to a person, and the crawler gets no help telling the service from the blog post from the footer.
Blocking “all AI” with one plugin toggle usually blocks the search crawlers along with the training ones. Check robots.txt for OAI-SearchBot, PerplexityBot and Google-Extended before deciding you were never going to show up anyway.
The fixes, in order
Cheapest and biggest first.
Let the right crawlers in. One line per bot in robots.txt, a rule in the WAF that treats them like Googlebot, and a test fetch with each user agent to prove it. Half the sites that “don’t show up in ChatGPT” show up a week after this.
# Search crawlers in, training crawlers your choice
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: GPTBot
Disallow: /Say it in words. Every service page gets one plain paragraph near the top: what the service is, who it is for, what it costs, how long it takes, where you work. The paragraph on this site’s WordPress speed page reads like a sentence a model can lift, because it was written to be lifted. Then the FAQ, with the questions buyers actually ask, as text on the page.
Repeat it in schema. Organization or ProfessionalService for the business, Service with an Offer for each service page, FAQPage for the questions, Article with an author for the posts. It is the same information in a form the parsers were built for, and it is a day of work on a site that already has the words.
Add the map. A text file at /llms.txt that lists what you offer, with prices and links. Not every assistant reads it yet. The ones that do get the whole offer in one request.
Stop sending a shell. If the site is client-rendered, server rendering or static pages for the marketing site is the fix. It is also the speed fix, which is not a coincidence.
- robots.txt allows OAI-SearchBot, PerplexityBot and Google-Extended
- A fetch with each crawler’s user agent returns the page, not a challenge or a 403
- The HTML contains the main text before any JavaScript runs
- Each service page says what, for whom, price and time in plain words
- Organization, Service and FAQPage schema validate without errors
- /llms.txt exists and matches the site
Six weeks later
The same twenty questions, the same three assistants, the same sheet. The comparison is the deliverable, because an AI answer on one day is not a result; the change between two days is. On the sites where the crawlers had been blocked, the second run is a different document. On the sites where the problem was vague copy, the change is slower and shows first in the by-name questions, because those are the ones that cite your own pages. Either way you have the two sheets, and you can see which questions still need a page written for them.
Related serviceWant the sixty answers for your own company?AI search readiness from $810: twenty buyer questions in three assistants, before and after; crawler access, service copy and schema fixed in between.See the serviceThe free scan already checks the first item on the list, crawler access, with the real user agents. It takes a few minutes and tells you whether the rest of the list is even reachable.
Questions from this article.
No, and anyone who does is selling something else. What I can do is make sure the answer it gives is right when it mentions you, that it can read your site at all, and that your service pages say in plain words what you do and for whom. Results are shown side by side, before and after, not promised.
Mostly. The same clear structure, the same schema, the same fast and readable pages help both. The differences are at the edges: AI crawlers have their own user agents and their own blocklists, they read text more than they read links, and a plain summary file helps them in a way it never helped Google.
Decide per bot. Blocking the training crawlers (GPTBot, CCBot) is a reasonable choice for a publisher. Blocking the search ones (OAI-SearchBot, PerplexityBot, Google-Extended for AI Overviews) means you are not in the answer when a buyer asks. Most B2B sites block both by accident, through a plugin default.
It is young and the big assistants have not promised to read it. It costs an hour, it is a clean map of what you offer, and this site has one. I treat it as cheap insurance, not as the main fix. The main fix is pages that say clearly what you do.