ChatGPT SEO: how to get your business cited in AI answers

Introduction

Someone asks us about the AI index every week. The one where you register your company. It does not exist, and the question behind it still deserves a longer answer than two sentences.

Start with a distinction that governs everything else. Google's AI features, meaning AI Overviews and AI Mode, draw on the same index as the classic ten results. Google describes the mechanism as retrieval augmented generation with a query fan-out in front of it: the model issues several related searches at once and builds its answer from the pages those searches return. ChatGPT is not a Google product, uses its own crawler and follows its own documentation.

What follows is a practical six-step checklist in this order, with concrete checks and a definition of when each step is finished.

Step 1: check access

This step comes first because everything after it depends on it. We have taken over sites whose copy had been polished for months and which the relevant crawler was never allowed to fetch.

Open yourdomain.com/robots.txt and read the whole file, not just the first group. Four entries matter: User-agent: *, Googlebot, OAI-SearchBot and GPTBot. OpenAI documents OAI-SearchBot as the crawler behind ChatGPT's search features, and a site that blocks it stops appearing in ChatGPT search answers. GPTBot governs model training and is a separate setting. Confusing the two locks you out of the answers when all you wanted was to opt out of training. A change takes roughly 24 hours to reach OpenAI's search systems.

Two special cases. ChatGPT-User handles fetches a person triggers, so robots.txt rules may not apply to it, and it does not decide whether you appear in search. Google-Extended controls training and grounding in other Google systems, not access to Search.

Then the page level. noindex removes the page, nosnippet stops any text from it being shown, data-nosnippet cuts individual paragraphs, max-snippet caps the length. To appear in Google's AI features a page has to be indexed and eligible to be shown with a snippet; Google lists no further technical requirements. The site also has to be included in Search generative AI features in Search Console.

That leaves the block that is not in robots.txt at all. Bot rules at Cloudflare or in a web application firewall answer unknown user agents with a 403 regardless of what you allowed. Check your server logs, filtered by user agent. And because a user-agent string can be set to anything, check hits against the IP ranges OpenAI publishes per bot.

Finished when: every target URL returns a 200 to Googlebot and OAI-SearchBot, Search Console reports it as indexed, and no nosnippet truncates the passage you care about.

Step 2: fix the entity in one sentence

"Think in entities" is advice nobody can act on. Write the sentence instead.

The pattern: [Legal name] is a [category] in [place] providing [service] for [audience].

Four rules go with it. The sentence stays under 25 words. The category word is the word your customers search with, not the one from your mission statement. Adjectives nobody can verify come out. And the place is the one in your legal notice, even if you work across borders.

Ours reads: Rankprofi is an SEO agency based in South Tyrol, working with local businesses and mid-sized companies across the DACH region and northern Italy.

The check that actually tests the sentence: give it to three people outside your company and ask each of them to name two competitors. If they name businesses from a different sector, your category word is wrong. That is the most common reason a model knows a company but does not suggest it when the question that matters comes up.

South Tyrol adds a wrinkle. The same company is Bozen in one source and Bolzano in another, and both spellings sit somewhere on the web. Pick one primary form per language version and mention the other once in your about text.

Finished when: the sentence lives in a file, carries a date and an owner, and becomes the template for every other description of your business.

Step 3: reconcile contradictions across sources

Now the sentence becomes the yardstick. Build a table with six data points in the columns: legal name, brand name, category, address, service list, founding year. Down the rows: home page, about page, legal notice, service pages, Google Business Profile, LinkedIn, chamber of commerce entry, directories and press mentions.

Four contradictions come up again and again. The legal name carries a GmbH or Srl in the legal notice that appears nowhere else. The location drifts between town, province and country. The category slips, because somebody typed "advertising agency" into a directory years ago and the website now says "SEO agency". And after a move, the old address lives on in three directories.

The order of repair is not arbitrary. Your own website first, because it is the source everyone else copies from. Then the profiles you administer; Google's documentation on AI features specifically asks you to keep Google Business Profile and Merchant Center up to date. Third-party entries last, because you can only request those and they take weeks.

Finished when: the sentence from step 2 appears word for word, or close to it, in every source you control, and no two of those sources disagree on the six data points.

Step 4: put the identity in machine-readable form

First the part most articles on this topic leave out. Google's own documentation says structured data is not required for generative AI features and that there is no special schema.org markup for them. We do not dispute that and add Organization markup anyway, for two reasons. It is the one place where the link between your site and your profiles is written down, and it makes the table from step 3 machine-checkable.

Concretely: one Organization object, identical site-wide, with name, legal name where it differs, URL, logo, address and contact point. Plus sameAs pointing at exactly the profiles from step 3. Google's Organization documentation also names vatID, iso6523Code and naics as properties that distinguish an organisation from identically named ones, which matters more than it sounds in a valley where four businesses share a family name.

The rule that gets broken most sits in Google's AI features guide: structured data has to match the visible text on the page. If your markup claims something the page never says, fix one of the two.

To check, run the Rich Results Test, then read the JSON-LD out loud against the visible page, line by line. The second part catches the errors the test misses, because those errors are formally valid. Setting markup up in general is covered in our article on setting up Organization markup.

Finished when: exactly one Organization object exists, no second version with different data is buried in the theme or the SEO plugin, and the sameAs list matches your table.

Step 5: write passages that stand alone

The inconvenient source again, up front. Google writes that you do not need to rewrite content for AI systems and do not need to break it into small pieces; chunking sits on the list of things you can ignore. The same guide carries the sentence worth working to: content people find unique and useful will influence your presence in generative AI search in the long run more than any other suggestion in it. Google's word for the opposite is commodity content, text built from common knowledge that could have come from anyone.

What we do with that is older than AI search: paragraphs a journalist could quote without calling first.

The test takes 30 seconds. Copy one paragraph into a blank document and read it. If it needs the heading, the previous sentence or an "as mentioned above" to make sense, it failed.

Five rules help. The first sentence names the subject rather than pointing back at it with "this" or "that approach". One paragraph carries one claim. Numbers arrive with a unit, a period and a source in the same sentence. References like "here" and "see below" come out. And important content sits there as text, not only inside a graphic, which Google lists too. Our house rule for length is 40 to 70 words: below that the condition that makes the claim true goes missing, above it a second claim creeps in.

Whether this raises the odds of being cited in ChatGPT is something no provider documents, so we do not claim it.

Finished when: every H2 section contains at least one paragraph that survives the cut-out test.

Step 6: measure without lying to yourself

On the Google side there is real data, in two places. The generative AI performance report in Search Console shows how your content does there, and Google counts AI Overviews and AI Mode in the ordinary performance report under the Web search type. Anyone who knows only one of the two reports thinks their AI data is smaller, or more complete, than it is.

For ChatGPT no equivalent exists. You measure by asking, which makes the protocol matter more than the result.

This is how we build the prompt set: 10 to 20 prompts in the customer's language, phrased the way a customer phrases things, in four groups. Category plus region. Problem description without the category word. Comparison question. And the direct question about the company.

For each prompt, log the date, platform, visible model version, whether you were named, where in the answer, which URL was cited and which competitors showed up. Always in a fresh session with no history, always on the same day of the month.

Tell your board this before the first number lands on the table: answers vary between sessions, users, regions and model versions. The number is a trend line, not a position. Google also warns about tools that promise ranking success or claim internal Google metrics; none of them has access to those systems. The full AI visibility protocol is in our article on measuring share of voice in ChatGPT and Perplexity.

Finished when: at least two measurement points exist before anyone presents a conclusion.

What we deliberately skip

We do not build a strategy on llms.txt. Google writes that its own Search ignores such files and that having one neither helps nor harms visibility there.

We do not buy mentions or pay for slots in ranking lists. Google calls seeking inauthentic mentions unhelpful and points at its own spam policies. Bought mentions also leave a link profile that costs more to clean up later than the mentions cost in the first place.

We do not build a page per fan-out query: producing content at scale to influence rankings or AI answers falls under Google's scaled content abuse policy. The term LLM SEO currently gets used for exactly that, which is why we rarely use it.

And we guarantee no mention, because Google states that indexing and serving are never guaranteed even when a page meets every requirement.

What this looks like on the Google side is covered in how AI Overviews rewrite the first page of Google. If you would rather not work through the procedure yourself, you will find it with us under visibility in AI answers; the free analysis is a starting point.