Your client isn't in the AI answer because they're not on the pages it read

You know the call. A client types their own category into ChatGPT — half curiosity, half ego — and up comes a tidy paragraph naming four companies. Three are competitors. One is a competitor they don't rate. They are not in it.

And now it's your problem, on a Tuesday, with no budget attached.

They're not in the answer because they're not on the pages the engine read to write it — and those pages are almost never their website. I put fifty real buyer questions to five engines about ten national brands and then went and opened 2,357 of the pages behind the answers. Reddit came back most, three times more than anything else. Then YouTube, Amazon, Forbes, Walmart, Trustpilot. Not one brand's own site made that list.

So the job isn't fixing their website. It might not even be a visibility problem — a couple of these causes are technical and take five minutes to rule out. Here's the order I'd work it in, cheapest and most binary first, because the causes are wildly different sizes. One you can settle before your coffee goes cold. Another takes eighteen months.

Everyone starts by auditing the site. Start one step earlier, because on the engine your client just tested, there may have been nothing to find them in.

ChatGPT went out to the live web on 166 of its 500 answers. Two in three came from memory. And it quit hardest exactly where buyers decide — on "X versus Y" questions it searched 4 times out of 80.

That matters because it means your client's best page cannot win a search that never happened. The other engines do look — Perplexity searched every time, Gemini nearly always, Claude most of the time — so whether an engine goes and looks is a property of the engine, not of your client, their fame, or their content.

So before you diagnose anything, note which engine they tested. If it was ChatGPT on a comparison question, you may be looking at the engine's habit rather than their absence. Run the same question on Perplexity and see whether the picture changes.

This is the five-minute one, and it's free, and it makes everything else pointless if you get it wrong.

When I sent a reader after every cited page in the study, the group that failed hardest was the one nobody guesses. More than half of the brands' own cited pages wouldn't open — and most of those refusals were the brand's own bot protection turning the reader away.

It's a setting, not physics. Same day, same reader: OLIPOP's site opened on all fifteen of its cited pages. AG1's blocked all twenty-two. One of those companies decided something; the other let a security vendor's default decide it, which is what almost all of these turn out to be — switched on by somebody in IT protecting the site from scrapers, because that is exactly what an AI crawler looks like from the inside.

Run this before you promise anybody anything:

curl -I -A "PerplexityBot" yourclientsdomain.com

Read the status code. Then check the bot-management layer separately from robots.txt, and check each crawler by name — one rule rarely covers them all.

The next fast one is rendering. A lot of AI crawlers don't execute JavaScript, so a site that assembles its content in the browser can be online, beautifully designed, and functionally blank to the thing you're trying to impress.

Fetch the page without JavaScript and look at what's actually in the HTML. If it's a loading skeleton and a div, you've found a one-to-four week engineering conversation — and you've found it before you sold anybody a content programme that could never have worked.

"We're invisible in AI" is almost never true across the board, because the engines don't read the same web.

Take Reddit, the single most-cited site in the study. For Perplexity it's about a seventh of everything it cites. For Google's AI Overview, near an eighth. For Gemini, four percent. For ChatGPT, two citations in a thousand. And Claude cited Reddit exactly zero times in 7,260 citations.

Same questions, same brands, same day. So a client who lives in the forums can be strong on Perplexity and absent from Claude, and the fix for one is not the fix for the other. A source strategy built by looking at one engine's citations is a strategy for a fifth of the market.

Test across all of them before you write a scope. Otherwise you'll sell a Reddit programme to a client whose buyers are on the one engine that has never cited Reddit in its life.

Once the door's unlocked and the content's readable, you run out of things you control. This is the real cause for most brands, and it's the slowest to fix.

Of the pages I read and classified, the biggest single slice is somebody's ranked list — "best cookware brands," "top ten coolers," your client sitting at position four in a round-up they didn't write. Another fifth is forum threads. Nearly a fifth is video. Actual editorial, a journalist writing about the space, is about one page in seven.

Read that list as an SEO and the problem is obvious. You can't add schema to a YouTube review somebody else filmed. You can't edit the round-up that ranked them fourth. A model doesn't decide your client is credible by reading their homepage — it decides by noticing that other places already treat them as a known thing.

Which reframes the ask. This is digital PR, community presence and review-platform work, and if your shop doesn't do PR, notice that the job just changed shape before you quote for it. Gareth Hoyle's twelve-causes breakdown is the best short list of the rest, and it comes with a line worth putting on a wall: most brands have two or three of these in play at once.

Here's the bit that only shows up if you actually read the pages rather than count them, and it changes what "we're mentioned" is worth.

A citation count says a page mentions your client. That can mean the whole page is about them, or one line in a list of twenty-five. Three real entries from the record: Gymshark appearing as item sixteen of twenty, roughly 5% of the page. Caraway holding first, third and fourth in a round-up of twenty-five products. And Caraway again — this time in a section headed "Nonstick pans we don't recommend."

All three are a mention. Only one of them is worth anything, and one of them is actively costing money.

So when you report on this, report where they appear on the page and what it says, not how many times a domain came back. A client who is technically cited in twelve places and buried at the bottom of all twelve does not have a visibility problem to celebrate.

Set expectations on these early, because otherwise they get charged to you.

A generic brand name. If the client is called something that's also an ordinary English word, the model has to guess which one you mean and it guesses wrong. Disambiguation is real work, six to eighteen months of it.

Missing from the structured sources. Wikipedia and its relatives. Three to twelve months, and partly outside anyone's control.

Negative framing. If the brand is present but discussed badly, you're not building visibility, you're shifting a narrative. Different job, different budget. And it's more common than people expect — half the forum threads the engines cite are more than two years old, so a complaint your client fixed in 2024 is still out there answering questions on their behalf.

The model predates the client. If the company didn't exist when the thing was trained, nothing lands until the next model does. You can't sell a fix for that. You can only explain it, which clients appreciate more than you'd think.

Plenty of "we're invisible" panics turn out to be a client testing one query they care about — usually the vain one, their own name, or a category label nobody outside the company uses.

Their buyers ask different questions. Test what a buyer would actually type, across every engine, more than once. Sometimes the finding is that the gap is real but somewhere else entirely, and that's a far better meeting than the one where you spend six months fixing the wrong thing.

If it helps to show a client what a finished diagnosis looks like, we published ten audits of national brands. Here's an example.

Run the curl command against their domain and see whether their front door even opens. Then search their category on Reddit and read what comes back — not the ranking, the conversation — and count how many times a competitor gets named and your client doesn't.

That's an afternoon, it costs nothing, and it tells you which half of this you're dealing with before you write a single line of scope.

One honest note to finish on. The study settles where the citations come from, how old they are, and how rarely a brand's own site is in the picture. What it doesn't settle is which of these causes is hurting your client most — nobody has that number, me included. Treat the order above as a sensible way to spend money, cheapest and most binary first, rather than as science. And be suspicious of anyone selling you a percentage with no method behind it. There's a lot of that about.

Why isn't my brand showing up in ChatGPT?

Usually two or three reasons at once, and the first one isn't about your site. In StyleForge's field study of five AI engines answering fifty buyer questions about ten national brands, ChatGPT went out to the live web on only 166 of its 500 answers — and just 4 times out of 80 on "X vs Y" comparisons. When it does search, it reads third-party pages: Reddit, YouTube, Amazon, Forbes, Walmart and Trustpilot topped the list of 21,213 citations, and no brand's own website made it. The fast technical causes are blocked AI crawlers and JavaScript-rendered content; the slow and most common one is that not enough independent pages discuss the brand.

How do I check whether AI crawlers can read my site?

Run curl -I -A "PerplexityBot" yourdomain.com and read the status code, then check your bot-management layer separately from robots.txt. It matters more than people expect: in StyleForge's field study, 53.8% of the brands' own cited pages wouldn't open for a machine reader, and 70 of those failures were the brand's own bot protection. It's a setting, not a limitation — OLIPOP's site opened on all 15 of its cited pages the same day AG1's blocked all 22.

Why does my brand show up in one AI engine but not another?

Because the engines don't read the same web. In StyleForge's study of 21,213 citations, Reddit made up 14.9% of Perplexity's citations and 11.8% of Google's AI Overview, but 0.2% of ChatGPT's and exactly zero of Claude's 7,260. A brand strong in the forums can be visible on one engine and absent from another, so test all of them before scoping any work.

Do backlinks help with AI visibility?

Less than you'd expect. The pages AI engines actually cite are mostly ones you can't link-build your way onto — in StyleForge's field study, 27.8% of cited pages were ranked lists, 22.7% forum threads and 19.7% video, against 14.9% editorial. Practitioners consistently report that brand mentions and branded search track AI citation more closely than backlink counts or Domain Rating.

Is schema markup worth doing?

Yes, and it's cheap. Organization, Article, FAQPage and Product schema remove ambiguity about who's behind a page — which matters most for brands whose name is also a common word. It won't put you on somebody else's round-up, which is where most AI answers are actually sourced.

Can I do anything if the model was trained before my company existed?

Not directly. You build the third-party presence now so the next model has something to learn from, and you make sure the live-retrieval surfaces can reach you today. Anyone promising more than that is selling.

• Part Two · The Source Study — every page behind the answers, what it is, how old it is, and whether it opens.
• GEO vs SEO: what the hell is the difference? — why the pages that decide this aren't yours.
• How to make money with AI visibility — what to sell once you've found the gap, and who'll pay for it.
• Citation building for AI search — the third-party presence work, in practical order.