Get Found in ChatGPT: The 90-Day Plan

1. Write down 15 questions your buyers actually ask

Do not start with your company name. Start with the decision your customer is trying to make. Write three questions for each of five types: the best solution for a specific case, the comparison of two routes, switching away from the incumbent, the cost, and the typical mistakes when buying or implementing.

A useful question is specific. "What is the best CRM?" shows nothing. "What is the best CRM for a 12-person recruiting agency that needs email sequencing and has nobody to maintain the system?" shows which vendors the AI considers and which evidence it looks for.

In GEO Tool: the workspace proposes questions by type (best, comparison, alternative, pricing, problem, how to choose, implementation) and uses Search Console queries as a starting point. Questions about typical mistakes you write yourself.

2. Measure who ChatGPT trusts today

Run each question three times in a fresh chat with web search on. Answers vary, so a single run is not a baseline. Log for each run: which brands are recommended, which URLs are cited, what kind of page each source is, and which sources recur.

After 45 runs, patterns become visible. For "best" questions the AI often cites no vendors at all, but reviews, comparison pages, trade publications, forums or videos. That list is your distribution plan: not "high-authority sites", but the exact sources already appearing for your buyers' questions.

In GEO Tool: the workspace runs the questions regularly against ChatGPT, Gemini, Perplexity and Google AI Overviews and stores for each run whether you are named, at which position, which competitors appear and which URLs are cited. The "Sources & Authority" view groups cited pages by type and shows which recur across questions.

3. Make your company impossible to misread

Before investing in new content, remove the contradictions. Homepage, About page, product pages, pricing page and third-party profiles should agree on what the company is called, which category it belongs to, what the product does, who it is for, where it is available and who is behind it.

Add structured data for organization, product, article and author where it fits. Keep the facts in visible copy too; schema supports the page and must not contradict it.

Check crawler access. OpenAI describes OAI-SearchBot as the crawler used to surface websites in ChatGPT search. GPTBot is for model training and is separate. A site that blocks OAI-SearchBot cannot appear in ChatGPT search.

In GEO Tool: the free GEO Check tests per page whether an organization or person with linked profiles is declared, whether the schema is valid, whether an author is named and which AI crawlers robots.txt blocks. Whether facts agree across several pages and profiles is something you check yourself.

4. Build one asset worth citing

Most companies publish opinions. AI answers need evidence. Pick one format: an original benchmark from anonymized product or customer data, a teardown of 50 to 100 websites or workflows, a transparent comparison with criteria and limits, a calculator built on real assumptions, or an analysis of customer results with sample size and method.

The asset needs one sentence another writer can safely repeat: sample, method, result. An example of the form: SE Ranking analysed about 129,000 domains and reports that domains with up to 2,500 referring domains averaged 1.6 to 1.8 ChatGPT citations, while domains with more than 350,000 averaged 8.4. One defensible number travels farther than thirty generic articles.

In GEO Tool: you collect the data yourself. The free Data Asset Builder calculates sample, average and comparison from it, writes the sentence for the top of the page, the table and the method, and checks that your text contains no number that does not come from your data. It runs in the browser without a language model.

5. Do not hide the finding

Do not make the reader dig the result out of 1,800 words of setup. Directly below the headline belong the answer in one sentence, the most important number, a table or chart, and a sentence on how the data was collected.

Then answer the obvious follow-ups under literal headings: What does it cost? Who is this for? How was this calculated? Clear writing is not dumbing down. It removes the reasons a machine might misread you.

In GEO Tool: the GEO Check tests for the answer in the first paragraph, tables, headings phrased as questions and evidence in the text. In the workspace the same check runs against the individual buyer question.

6. Own one specific buyer problem

Semrush tracked 1,094 categories in ChatGPT over six months. Only 15.2% had a clear leader in the answers, 31.2% an emerging one, and 53.7% none at all. In most categories the position is still open.

Do not try to own "project management software". Own a smaller decision, such as "project management software for architecture firms with client approvals". Build five pages around it: the complete guide, the best options and who each suits, the comparison with the incumbent, the pricing and total cost, and the implementation checklist. Each page answers a different question, all use the same terms and link to each other.

In GEO Tool: the workspace creates a brief per question. The five pages as a connected set, and the links between them, you plan yourself.

7. Get mentioned where your buyers already learn

Ahrefs studied about 75,000 brands. For ChatGPT, visibility correlated most strongly with mentions on YouTube (0.737), then with branded web mentions (0.664). Domain Rating was at 0.266. Correlation is not causation. The gap is still hard to ignore: being talked about tracks AI visibility more closely than one authority score.

Take your citation map and choose 25 targets: ten trade publications, newsletters or resource pages, five creators, podcasts or YouTube channels, five comparison, directory or review pages, and five customers, partners or associations with a website of their own.

Do not ask for "a link". Offer the finding, the data behind it, a chart they can embed and a sentence on where the result does not apply. The pitch is not "please mention us" but "we found something your readers will want to know".

In GEO Tool: the workspace helps you choose a target source, pick the fitting route and prepare the pitch. Nothing is sent automatically; you make the contact.

8. Fix the review profiles that already appear in answers

In the SE Ranking analysis, domains absent from the major review platforms averaged 1.8 citations, while domains listed on several platforms averaged 4.6 to 6.3. This too is an association, not proof that four new profiles create citations. It shows why third-party corroboration deserves attention.

Choose the two or three platforms that already appear in your citation map. Claim the profile, complete it, use the same category and description as on your site, and ask real customers for honest reviews. Manufactured reviews are not a shortcut.

In GEO Tool: review platforms show up in the citation map when the AI cites them. Whether your profile there is complete is not something the tool checks.

9. Track whether citations create enquiries

A screenshot saying "we appear in ChatGPT" is not a metric. Track five things: the share of your 45 runs that cite you; the share that name you; which third-party sources recur; the visits that come from ChatGPT; and what those visits turn into.

You can recognise visits from ChatGPT in web analytics by the parameter utm_source=chatgpt.com, which ChatGPT appends to linked URLs. In our own measurements 741 of 743 URLs cited by ChatGPT carried it.

Keep the question list stable for the whole period. If you rewrite it every week, you are changing the test instead of measuring progress.

In GEO Tool: the workspace shows per question whether you are named and whether your page is cited, and compares measurement runs before and after a change. With Google Analytics connected it shows visits from AI systems. Enquiries and revenue from those visits you evaluate in your own analytics.

What is realistic after 90 days

You have a baseline and a comparison for the same 15 questions, one page with original data, one topic you own and a handful of mentions. Whether that turns into citations is not decided by this plan, but by whether the finding is useful to others.

You can still hire an agency afterwards. But you will be able to tell whether it builds real evidence. The goal is not to trick ChatGPT. The goal is to be the easiest credible answer on the web to cite.