I keep getting a version of the same question from squirrelscan users: "our SEO is fine, so why doesn't ChatGPT ever mention us?" The honest answer is that "SEO is fine" and "ChatGPT can find and cite us" are two different systems with some overlap, and most people are optimizing for a search algorithm that isn't the one deciding whether they get cited. There's also a second, entirely different search intent hiding behind the same keyword: people who type "chatgpt seo" because they want ChatGPT to do their SEO work for them, not because they want to be found by it. Both deserve a straight answer, so here's how the actual mechanics work, then what to do about each.
The three crawlers behind ChatGPT
OpenAI runs three separate bots, documented in its overview of OpenAI crawlers, and treating them as one thing is the single most common mistake we see in the audits we run. GPTBot only collects training data for future models, so blocking it just opts you out of training and nothing else. OAI-SearchBot is the one that actually matters for citations: it pre-crawls and indexes pages to surface them in ChatGPT's search feature, and blocking it drops you out of search answers, though you may still turn up as a bare navigational link. ChatGPT-User works differently again. It fires on demand, the moment a live user prompt triggers a browsing action, like asking ChatGPT to open a specific page. That's a real-time fetch, not a background crawl, and robots.txt rules may not even apply to it the same way since a human explicitly asked for that page.
Each one is gated independently in robots.txt, so you can allow OAI-SearchBot for search visibility while still blocking GPTBot from training on your content. A lot of sites get this backwards: they add a blanket "block all AI bots" rule after reading a headline about training data, and accidentally take themselves out of ChatGPT's actual search answers in the process.
Where the index behind it actually comes from
This is the part nobody outside OpenAI can answer with full confidence, and I'd rather say that plainly than make something up. ChatGPT's search feature launched running on top of Bing's index and search infrastructure. Since then, OpenAI has talked about building its own crawl-based index, largely powered by OAI-SearchBot itself, and independent researchers poking at citation patterns have found evidence that looks like more than one backend depending on the query. Bing's own guidance on how its index and ranking work is still a reasonable proxy for baseline crawlability, since a page that can't get indexed there is unlikely to fare better anywhere else.
The practical takeaway: don't chase a specific "ChatGPT algorithm" the way you'd chase Google's. Optimize for the things that make any of these systems, present or future backend, able to fetch, parse, and quote your page. That's a shorter and more durable list than it sounds.
What actually gets cited
Once a crawler can reach a page, three things determine whether it's the one that gets quoted back to a user.
Quotable structure. ChatGPT's answers lift specific passages, not whole pages. A section that opens with the direct answer to the question it's about, in the first sentence, survives being extracted far better than one that builds up to the point over three paragraphs.
Freshness that's visible, not just true. A page that was updated last month but doesn't say so anywhere reads the same to a model as a page that's been stale for two years. Put a visible date on anything time-sensitive.
Entity clarity. The model needs to know what you are before it can decide whether to cite you for a given question. Consistent naming across your site, an Organization or Product schema block, and a Wikipedia or Wikidata presence if you're a real brand all reduce ambiguity. Structured data doesn't guarantee a citation, but Google's own guidance on product markup is a good baseline even outside Google's own products, since the same schema gets reused by other systems, including ChatGPT's shopping surfaces for merchants on catalogs like Shopify's.
The checklist
- Confirm OAI-SearchBot isn't blocked. Check
robots.txtagainst the syntax defined in the original robots exclusion spec. If you have a catch-all bot block, carve out an explicit allow forOAI-SearchBot. - Lead every section with its answer. Rewrite the first sentence under each heading so it stands alone if lifted out of context.
- Keep an honest, visible "updated" date on anything you want considered current.
- Add structured data for what the page actually is: Article, Product, FAQPage, whichever matches.
- Publish an llms.txt file. It's not confirmed to be read by OpenAI's own crawlers, but it costs almost nothing and helps the growing set of agents and MCP tools that do fetch it.
- Make sure your main content renders without JavaScript. Several crawlers, OAI-SearchBot included, fetch raw HTML rather than running your client-side scripts.
Every one of these is a specific, checkable thing, not a vibe, which is exactly what makes it auditable instead of aspirational.
Using ChatGPT to do SEO work
ChatGPT is also a genuinely useful brainstorming and writing partner for SEO work: drafting title tags, outlining content, explaining a concept. What it can't do from a chat window is tell you what's actually broken on your site, because it has no live crawl of your pages, no real Core Web Vitals data, and no way to know your actual meta tags unless you paste them in one at a time. Asking ChatGPT to "audit my site for SEO issues" gets you a plausible-sounding list built from general knowledge, not a finding grounded in your actual HTML.
The better version of that same workflow is giving your coding agent real audit data instead of asking it to guess. That's the whole idea behind running squirrelscan from inside an agent: it crawls the real page, checks it against 249+ rules, and hands back structured findings your agent (ChatGPT included, once it's connected to real tools) can act on directly, instead of a hallucinated punch list.
Check where your site actually stands
The fastest way to know whether OAI-SearchBot can reach you, and whether the rest of the checklist above is in place, is to run an actual crawl instead of guessing from a chat window:
curl -fsSL https://install.squirrelscan.com | sh
squirrel audit https://example.com --format llmThe --format llm flag renders the report as clean text an agent can act on directly, which is the same principle as everything above: structure your output, and your input, for whichever reader is actually going to consume it.
For the wider picture beyond ChatGPT specifically, our generative engine optimization guide covers how Anthropic, Google, and Perplexity each handle discovery and citation differently, and our AI crawler guide goes deeper on robots.txt configuration across all of them. If you're setting squirrelscan up as a tool ChatGPT itself can call, see our ChatGPT setup guide.
Want to know exactly where your site stands with AI crawlers before you touch a single heading? Run a free llms.txt check, or grab the CLI and run a full audit locally, free, no account required.