Showing up in AI search is two contests, not one. First the engine retrieves a set of candidate pages for the question; then it decides which two or three to name in the answer. A site can win the first and lose the second every time, and nothing in Google Search Console will tell you. If you are not retrieved at all, the problem is discovery — crawler access, thin content, missing schema. If you are retrieved and never cited, the problem is quotability, and the usual cause is that your one useful specific is buried four paragraphs down.
In three ChatGPT answers about local SEO pricing recorded on 2026-08-12, the citations went to Reddit, BrightLocal, Search Engine Land, SEO.com and mediasearchgroup.com. Every one of those pages opened with an explicit dollar range or a price-tier table in the first screen. The pages that were retrieved and dropped, mine included, opened with context and reached the number several paragraphs later. Position of the specific, not quality of the prose, separated the two groups.
What the pull actually showed
On 2026-08-12 I ran DataForSEO's LLM-mentions endpoint against my own domain to stop guessing. It came back with three ChatGPT prompts: whether local SEO is worth it, how much local SEO costs, and how much local SEO services cost. Combined, those three carry about 255 AI searches a month.
My guide on local SEO pricing appeared in the retrieved search results for all three. It appeared in the cited sources for none of them. The citations went to Search Engine Land, BrightLocal, SEO.com, mediasearchgroup.com — and Reddit, twice.
I had assumed I was simply not in ChatGPT's index for those questions. I was in it every single time and being passed over at the last step, which is a considerably more annoying and considerably more fixable problem.
Why the pages that won, won
I read them. There is no clever trick in there. Every page that took a citation put a number where a model composing an answer would hit it immediately: SEO.com opens with "$500-$2,000 per month", mediasearchgroup.com with "between $300 and $9,500 per month in 2026", hellowebdesigns.com with a "Quick Answer" block before anything else. Two of the three ChatGPT answers rendered as a price-tier table, and the sources it drew on were the ones already shaped like a table.
Reddit is the outlier worth noticing, because it wins on something you cannot copy directly: a thread of named people saying what they actually paid. That is first-hand pricing testimony, and an engine treats it as a distinct kind of evidence from a vendor's rate card. The nearest honest equivalent for a small business is publishing your own real numbers rather than a range you got from a survey.
My page reached the same numbers. It just did so after establishing context, the way you would if you expected a human to read top to bottom. Assistants do not read top to bottom. They lift.
Diagnose which problem you have first
The fixes for the two failures are close to opposite, so establish which one you are looking at before changing anything. Pull the mention data for your domain and look at the two arrays separately: the sources the answer cited, and the search results it drew from. If your domain is absent from both, you have a discovery problem. If it is in the second and not the first, you have a quotability problem.
Discovery problems are technical and usually dull: AI crawlers blocked in robots.txt, no schema, content too thin to be a candidate. I found one of these on my own site the same week — the named crawler groups in my robots.txt were missing their disallow list, which is the opposite failure but the same lesson about checking rather than assuming.
Quotability problems are editorial. Move the specific to the opening block, put it in a table if it is a set of numbers, and make sure the sentence containing it can stand alone with no antecedent to resolve. A model quoting "it typically runs $500 to $3,000 a month" needs to know what "it" is without your previous paragraph.
What I changed, and what I am still waiting on
The honest state of this: I have published the baseline and rewritten the openings, and I do not yet have an after number. Citation patterns take time to move because an engine has to re-crawl, re-evaluate, and change a habit it has already formed. Anyone promising you a fortnight's turnaround on this has not measured it either.
What I can commit to is publishing the re-pull against the same three prompts, whatever it says. Zero of three is on the record. If it stays at zero of three, that will be on the record too — which is more than most of the pages that beat me will ever tell you about their own results.
Key takeaways
- Retrieval and citation are separate outcomes; only the second one sends you anyone.
- Pull the mention data before changing anything — the two failures need opposite fixes.
- Every page that beat me led with an explicit number or a price table in the first screen.
- Make the sentence containing your specific stand alone, with no antecedent to resolve.
- My baseline is 0 of 3 prompts as of 2026-08-12, and I will publish the re-pull.
Common questions
Pull recorded mention data for your domain rather than prompting ChatGPT yourself — your own prompting is a sample of one and answers vary between sessions. DataForSEO's LLM-mentions endpoint returns the prompts where a domain has actually appeared, split into the sources an answer cited and the search results it drew from. If you are in neither array, start with crawler access and schema. If you are in the second only, your content is being read and skipped.
Because a Reddit thread on pricing is a set of named people reporting what they actually paid, and that is a different category of evidence from a vendor publishing its own rate card. An engine synthesising an answer about cost has an obvious reason to prefer it. You cannot copy that directly, but you can get closer to it by publishing your own real numbers — what you charge, what a real project cost — instead of a range borrowed from an industry survey.
Usually yes, as a side effect. Answer-first writing with the specific up top is the same structure that wins featured snippets, so the work overlaps. But do not do it for that reason. Do it because the two channels are diverging, and a page can hold a decent organic ranking while quietly losing every AI citation attached to the same query.
I do not know yet, and I would be suspicious of anyone who claims a confident number. The engine has to re-crawl the page, re-evaluate it against the prompt, and displace a source it has already settled on — and that last part is the slow one. My own baseline of zero citations across three prompts is dated 2026-08-12 and published here specifically so the answer is verifiable later rather than asserted now.
Related guides
Want this done for your business?
I build fast, schema-rich websites for Massachusetts service businesses — engineered for local and AI search from the first line of code.