GEO and SEO: how to check your visibility in AI answers

GEO, or Generative Engine Optimization, is the work of making a business and its content more useful in the answers AI assistants give. For a small business, start with a commercial question: when a prospect compares options, what do they find out about your offer, and is it accurate?

SEO still matters. Google confirms that its usual practices also apply to AI Overviews and AI Mode. You need no AI file or special markup to appear there.

Where does an assistant’s answer come from?

An assistant draws on two sources, often in the same conversation. The model learned part of its knowledge during training, from text collected months before its release. It can also run a web search when the question comes in, read a few pages and cite them as links under its answer.

Live search goes through identified bots. ChatGPT uses OAI-SearchBot for its search features, Perplexity uses PerplexityBot, Claude uses Claude-SearchBot, and Google’s AI Overviews rely on Googlebot. A page these bots cannot read has little chance of being cited in this mode. The robots.txt guide covers access bot by bot.

Training knowledge moves more slowly. If an old price or a discontinued service is in there, your correction will only show up in a future model, and no vendor publishes when that will be. The first effects of your corrections you can observe will therefore be in answers that use live search.

The term GEO comes from a research paper published in 2023 by Pranjal Aggarwal and his co-authors. They tested several content changes on a benchmark of queries and found that how well each one worked varied by domain. Keep that variation in mind when a provider sells you a method: it may have worked in a different sector from yours.

Which questions should you check?

Go back to the questions you hear in sales meetings: price, compatibility, choosing a provider, alternatives and delivery terms. Drop the queries nobody asks before buying from you.

Build a small, stable panel, ten questions for example. Include some phrasings without your brand name and others that ask for a specific fact about you. That split lets you tell whether assistants discover your business apart from whether they get your details right.

For each test, note the question, the assistant, the date, the search mode and the links cited. Replay the same questions in new conversations. A single answer can vary and does not represent your whole audience.

A grid like this one is enough, filled in at each pass:

Question	Assistant and mode	Date	Brand cited	Page cited	Wrong information
“Which provider for [your service] in [your city]?”
“Does [your brand] deliver to France?”

The last column counts as much as the others. A mention with an out-of-date price costs you more than no mention at all, because the prospect arrives with the wrong expectation.

What should you fix when an assistant describes your offer wrongly?

Check your own site first. A services page should explain the need it addresses, the deliverable, the limits, the terms and the budget if it is public. Prices and availability must match from one page to the next.

On an article, look for the precise answer the reader expects: a calculation, a procedure, a comparison or a compatibility limit. A long definition of the topic does nothing for a prospect who is trying to choose.

Link that answer to its evidence: the supplier’s documentation, an official source or a case study someone can check. The Nallian redesign, for example, describes work that was done. It does not prove that every project will get the same result.

When the answer cites a link, open it before you change anything. The error may come from an old page that is still indexed, a directory listing nobody updated or a dated press article. Fix it at the source: redirect the old page to the current one, update the listing, write to the article’s publisher. Then replay the same question once the bots have been back. Your server logs show the date of their visit.

Can people identify your brand elsewhere?

The same name must point to the same business on your site and on the profiles your clients look at. Check the activity, contact details, area served and links. Correct the old descriptions on existing profiles first.

Before you touch the profiles, write an internal identity sheet: exact name, short description, list of services, area served, site address and main pages. Anyone who updates a profile then works from the same text. Name variants such as “Dupont Conseil”, “Dupont Conseil SRL” and “DC Conseil” become easy to spot and align.

Start with the profiles your clients actually open. For a Belgian SME, look first at the Google Business Profile, the company LinkedIn page and the directories or review platforms in its sector. On your site, the sameAs property in structured data lists these official profiles. The Schema.org guide shows how.

Third-party sources can document a reference, a partnership or a review. Choose the ones that count in your sector. There is no universal list of platforms that guarantees an AI citation, and no reason to create a Wikipedia page for every small business.

Wikipedia and Wikidata have their own rules. Wikidata accepts an entry for a clearly identifiable entity described by serious, public references. French-language Wikipedia requires notability established by secondary sources, for example two articles in national media at least two years apart, and advises against autobiographies. A small business with no press coverage does not meet these criteria.

An interview about your trade or a client reference is worth the help it gives your prospects. Chasing mentions for their own sake tends to produce filler on other sites.

Which technical checks should you ask for?

• The important pages are accessible and can be indexed.
• Internal links lead to them from the relevant services.
• The useful text is in the page itself and stays readable on mobile.
• The structured data matches the visible information.
• The firewall and CDN let through the bots you want to allow.

The robots.txt guide explains access. The Schema.org guide covers data consistency. The llms.txt file can help some tools once these checks are done. If you also want an agent to be able to request a quote, the AI agents guide covers data, permissions and handing over to a person.

How do you judge results and budget?

Track correct mentions, links to your pages, identifiable visits and sales enquiries separately. An assistant can talk about you without sending a visit, and your analytics can also attribute a visit to the wrong source.

In your panel, record how many answers mention you and how many cite a useful page. Keep the same protocol so you can compare two periods. That rate describes your test, at your scale.

Google provides part of the measurement. Visits from AI Overviews and AI Mode are counted in the Search Console Performance report, along with the rest of your web traffic. Since 31 August 2026, Search Console also shows data on the pages that appear in these AI features, including impressions and countries (Google’s announcement). For ChatGPT, Perplexity or Claude, look at referral visits in your analytics tool and cross-check them with your panel.

Paid tools such as Otterly.ai, Profound or Peec AI automate this panel: they put your questions to several assistants and track mentions over time. Their output depends entirely on the questions you give them. A badly chosen panel gets you regular tracking of queries your prospects never ask.

Ask for a baseline, actions with a stated reason and regular measurement. A guarantee of being cited in every answer is a commitment nobody can verify.

Frequently asked questions

Does GEO replace SEO?

No. Google says the usual SEO practices apply to its AI features, and the live search used by other assistants also reads pages that are accessible and indexable. GEO adds two pieces of work to SEO: tracking what assistants answer and working on the third-party sources that talk about you.

Do I need a paid tool to start?

No. Ten questions replayed by hand each month and logged in a spreadsheet already give you a baseline and a trend. A tool becomes useful when the panel grows, or when you track several brands, languages or markets.

How long before I see an effect?

Nobody can give you a reliable timeframe. With live search, a corrected page can be picked up as soon as the bot next visits. OpenAI, for example, says a robots.txt change takes about 24 hours to affect its search results. For training knowledge, no vendor publishes the schedule of its next models.

Can a citation in ChatGPT or Perplexity be guaranteed?

No. Vendors do not publish how they select sources, and the same question gets different answers from one conversation to the next. A serious provider commits to a baseline, specific corrections and regular measurement with the same protocol.