Open ChatGPT or Perplexity and ask: “Which company should I choose for [your service] in [your city or country]?” Write down the companies mentioned, whether sources are linked and which facts the system uses to explain its choice. Ask the question again with slightly different wording. If your company is absent, that is not yet a diagnosis. It is a concrete starting point, though, and more useful than a presentation promising to “put you into AI answers”.
Generative Engine Optimization (GEO) is work that helps AI systems find, understand and use reliable information about a business when answering relevant questions. Much of the practical work will look familiar from SEO: accessible pages, clear services, useful explanations, consistent facts and independent corroboration. The newer question for an owner is how to observe whether the business appears in AI answers without pretending there is a “number-one position in ChatGPT”.
There is an important limit. A page, a technical file or one mention in one query cannot guarantee that a system will recommend you to everyone. Wording, time, available sources and product versions can change the answer. I would begin with a fixed set of questions and pages that can be checked, rather than a flattering screenshot from one successful chat.
That screenshot is fine as a note. It is a poor basis for a long-term channel budget.
1. How an AI answer finds a source
In ordinary search, people see a list of pages and open the ones they need. In Google AI features or ChatGPT Search, a system can assemble an answer with sources alongside it. That does not mean every model searches the web in every conversation: behaviour depends on the product and mode. The practical point for a business is simpler. If an important page cannot be opened, read or understood, a search system has less to work with.
Google Search Central explains that the usual SEO foundations apply to AI Overviews and AI Mode. For a page to be eligible as a supporting link, it must be indexed and eligible to appear with a snippet in Google Search. These features have no extra technical requirement or special schema.org markup. Google also cautions that eligibility does not guarantee crawling, indexing or a feature appearance.
For ChatGPT Search, OpenAI describes OAI-SearchBot as a crawler that helps surface sites in search answers. GPTBot serves a different purpose related to potential use of content for model training. You can set their access independently: allowing the search crawler does not decide what to do with GPTBot. It is an access decision, not a “get cited” button. Also check whether a site’s protection service blocks the relevant requests despite apparently correct robots.txt rules.
If your site cannot be reached by ordinary search, start with a technical SEO audit. An “AI file” will not help a page that returns an error or hides its main terms behind a login.
2. GEO and SEO: shared work, different observations
| Question | Conventional SEO | GEO / AI Search |
|---|---|---|
| What to prepare | Accessible pages, site structure and answers to search questions | The same pages, with clear facts a system can use as a source |
| What people see | Results, snippets and pages | An answer that may include links, sources or brand mentions |
| What to check technically | Crawling, indexing, internal links and visible text | The same, plus access for named search crawlers where a platform documents them |
| What to measure | Queries, impressions, clicks, landing pages and enquiries | A fixed sample of answers, mentions and sources, visits and quality of enquiries |
| What not to promise | A guaranteed ranking or traffic | A guaranteed recommendation or citation in every answer |
These are not two separate teams that need to rewrite the same page independently. A service page that honestly explains whom it serves, where the business works, what the process looks like and what its limits are is easier for people and search systems to assess. For an AI answer, it also helps to have specific facts that can be conveyed briefly and accurately, without contradictions elsewhere in public material.
Measurement needs care. According to Google’s guidance, visits from its AI features are included in the general web performance reporting in Search Console. This does not automatically give you a neat chart of “all AI Overview mentions”. Other AI systems show sources and analytics differently. A manual question panel, referral visits and real enquiries supplement search data; they do not replace it.
3. What to check on your website
1. The page is reachable and eligible for indexing. Check the real URL, server response, canonical, robots.txt, meta tags and whether other parts of the site link to it. Use URL Inspection in Search Console for Google; for other crawlers, consult the platform’s documentation and server logs when useful. Absence from a single AI answer is not evidence of deindexing.
2. Important facts appear as text. Company name, service, market, terms, process and contact details should not live only in an image. If the headline says “innovative solutions” while the actual service is buried in an unlinked PDF, both people and systems have to guess. Put answers to real buyer questions on the relevant page instead of hiding them across ten generic articles.
3. The facts agree across places. The website says “we work across Ukraine”, a company profile says “Kyiv only”, and an old article promises another country. Fix that discrepancy before discussing GEO. Independent mentions, reviews and profiles can help people verify a business when they are genuine. A purchased “top ten companies” article in which all ten entries belong to one owner does not become evidence because a crawler has seen it.
4. Markup matches visible content. Structured data can help a search engine understand the entities on a page, but it cannot replace the page text. Google explicitly warns that even correct markup cannot guarantee a special search appearance. Do not put invented awards, reviews or prices in JSON-LD. There is no separate “GEO schema” for Google’s AI features.
5. Buyers have a next step. A citation leading to a page without terms or a way to enquire may generate interest without helping the business. Look at what people see after the click. Our AI Search visibility checklist walks through technical and content questions on an existing site.
4. A real case: visibility index, not customer count
In GrandMa’s B2B AI Search case, we worked on visibility in Google AI Overviews, ChatGPT and Perplexity. The initial problem was that a strong product and conventional search work were not producing a presence in the selected AI answers. We looked at buyers’ questions and whether the public content explained the product in a way that could be used as a source.
The public case documents a fixed panel of four buyer questions and a normalised AI Visibility Index. Its baseline was set to 100. After five months, the index reached 420, a +320% change from baseline. This is a measure within the case’s defined question panel and method. It does not mean 420 citations, 420 customers, a ChatGPT ranking or a 320% sales increase.
Why name the method? If you ask one version of a question today and get a mention, then ask a different question tomorrow, you cannot tell whether anything improved. The fixed panel in the case allowed the index to be followed over five months. It still could not cover every possible market question. Such a measure indicates direction within its selected set; business outcomes need separate verification.
I would not use this B2B case figure as a forecast for a small business. It proves a result of a particular measurement in a particular project. The transferable principle is to select relevant questions, preserve the starting state, then improve pages and see whether their facts become easier to find and verify. We explain more in our B2B AI Search visibility article.
5. Ten questions for a monthly check
Replace bracketed text with your service, buyer and geography. Choose only questions your customers might genuinely ask. This is a self-check panel, separate from the four-question panel in the case. Do not ask each question once and call the share of mentions “AI market share”.
- “Which companies provide [service] for [buyer type] in [country/city]?”
- “How should I compare providers of [service] for [specific task]?”
- “Who can help with [specific problem] in [region] if I need [condition]?”
- “What should I ask a company before buying [service]?”
- “What are the limits of [category] services for [type of business]?”
- “What alternative to [common approach] is available for [task]?”
- “Where can I find examples of completed work on [topic] in [location]?”
- “Who provides [service] with [important condition: language, deadline, integration]?”
- “How does [your approach] differ from [other approach] for [task]?”
- “Which sources support the experience claimed by companies offering [service]?”
Remove “Recommend my company specifically”. It tests whether the model follows your instruction, not whether a buyer can discover you. Do not add ten near-identical variations to improve your mention rate. The panel should cover different decision moments: category discovery, comparison, checking terms and checking proof.
Table for recording the results
Ask the same questions in the same systems each month. Record the date, country, language and product mode. Start a new conversation without prior context where possible. Keep answers made with web search separate from answers made without it.
| Date and system | Question no. | Was our brand mentioned? | Link to us? | Who else appeared? | Which sources appeared? | Notes |
|---|---|---|---|---|---|---|
| [date], [ChatGPT Search / Google / Perplexity] | [1-10] | yes / no / unclear | URL or none | names | URLs and source types | language, region, factual error |
Check Search Console, server access logs where useful, referral visits and real enquiries separately. A form or CRM can ask how a person first heard about you, though self-reporting is not a perfect attribution report. If mentions rise but enquiries do not, check the linked page and the offer itself. If enquiries arrive, check whether they fit your service rather than merely counting them.
When an answer gets a fact about your business wrong
Do not begin by asking the model to “remember the correct price”. Locate the possible source: an old page on your site, an outdated profile, a third-party directory or conflicting service descriptions. Record the specific claim, source URL and date of review. Fix what you control. For another platform, use its official correction route. Then check the same fixed questions again, knowing that an updated answer may take time and no deadline is guaranteed.
If the system names a competitor, do not copy their page. Ask which buyer question the sources answered better. Perhaps your own site never states the areas you serve or shows an example for that type of customer. That is useful editorial work even if the next AI answer remains unpredictable. Accurate facts help a person verify your business through any discovery channel.
6. Warning signs in a provider's promise
“Guaranteed ChatGPT mentions in 30 days” has no honest way to verify the promise across all users and phrasings. “We’ll add a secret file and Google AI Overviews will pick you up” conflicts with Google’s explanation that these features require no special AI file or markup. “We can already see all your AI sales” without configured measurement is another reason to ask where the number came from.
A useful proposal is more specific: which pages will be checked, which facts clarified, what access issues fixed, where independent corroboration is needed, which questions will measure the work and what will remain unknown. If a provider cannot name the limits of its metric, the graph can be attractive and useless at the same time.
GEO should begin with work that helps even without an AI mention: people can understand the service, search systems can reach the pages and the team knows which claims it can substantiate. On GrandMa’s AI Search service page we explain this route through technical accessibility, content and measurement. A citation that follows is an observable additional outcome, not an entitlement to a guaranteed place in someone else’s answer.