AI search visibility (GEO)

Three things: a measurement, a fix list, and the fixes implemented. The measurement tells you how many answers out of three name your company today, and which competitors are named in your place. The fix list is ordered by priority and specific enough to hand straight to a developer. The implementation is that list built, either by us or by your own team.

The technical spine is eight checks, the same ones that run in the free checker on our homepage: whether the page opens for an agent, whether the title and description are present, whether structured data exists, whether robots.txt allows AI crawlers, whether llms.txt is published, whether prices are machine-readable, whether a machine can make contact, and whether an MCP interface exists. None of them asks a language model for a score, so the same page returns the same result tomorrow. The reasoning behind that approach is set out in what generative engine optimization is.

The retainer adds repetition. Every month the same blind test runs with the same questions, and the report carries the mention count, the competitors named in the same answer, and the work done that month. The report is short, and the number in it can go down. We do not change the questions to make the chart look better.

The test is blind, and the brand name never appears in the question. That is the single most important rule in the method. Ask a model what it thinks of company X and it will repeat what you fed it. The result looks like visibility from the outside, but it is a false positive, and no decision should be built on it.

The check runs in two phases. First we classify the business: what it actually does, where it operates, and which questions it should plausibly appear in. Then we ask those questions the way a customer would ask them, and count how many answers name the company. We also record who was named instead, because a competitor's name in the answer is more useful than your own absence.

Two results from our own measurements. aiasemu.ee, the lawn-care business we run ourselves and describe in how we run a business on AI, was named in two of three blind Gemini answers. The counter-example teaches more. We tested a holiday-let property whose classic search visibility is strong enough that strangers write to ask who did its SEO. Three blind questions about holiday houses in its region produced zero mentions out of three, and competitors were named in all three answers. Good Google visibility does not carry over to AI answers on its own.

First, robots.txt. Our own file allows twelve AI crawlers by name, GPTBot, ClaudeBot and PerplexityBot among them. The common finding on other sites is the reverse: a privacy or security plugin blocked them by default and the owner never knew. If the crawler cannot reach the page, the rest of the work is pointless.

Second, structured data and machine-readable prices. The most frequent finding in our measurements is a price locked inside an image or a downloadable PDF. A person can read it, a machine cannot, and when someone asks an assistant to compare suppliers your name is left out because the model has no number to compare. Schema.org markup for the organisation, its services and its prices gives the same text a meaning a machine can use, which is the gap behind why ChatGPT does not recommend your business.

Third, llms.txt and a contact path an agent can actually use. Our own llms.txt carries the full offer, the prices and the guide index. Our /api/lead endpoint answers a GET request with its own schema, so an agent that finds the URL learns the field format without guessing, then posts the enquiry. The same principle applies to your contact form.

Fourth, indexing. ChatGPT search and Copilot read mainly from Bing's index while Gemini leans on Google's, so we submit through IndexNow after every deploy. Roughly 7.8% of ChatGPT citations come from Wikipedia and Wikidata-style sources, which is the argument for keeping your company details consistent off your own site as well.

Audit, 290€, one-off. Your site goes through the eight checks point by point, the blind test runs, and you receive a prioritised list: what is broken, why it matters to a machine reading the page, and the order to fix it in. The fee is credited against follow-up work if you decide to have the fixes built.

Fixes, 990€, on an existing site. The list implemented: structured data, llms.txt, robots.txt, prices as text rather than images, and a contact path a machine can complete. The work takes about a week and closes with a repeat measurement, so before and after are comparable on the same scale.

Visibility retainer, 590€ a month. Ongoing content and technical work plus a monthly report on how often assistants name you. In practice that means building pages for the questions customers actually ask, and keeping the structured data correct as the site changes. There is no fixed term.

If there is no site yet, or the current one is being replaced anyway, a new build starts at 3 999€ per project and readiness goes into the brief from the start. We publish these prices so a buyer can decide without waiting for a quote. That habit is also part of the service itself: when a model is asked to compare suppliers, it can only name the ones that stated a number.

The audit needs a domain and nothing else. The fix work needs four things. Access to the site, usually the CMS, and to the server or DNS where necessary, because robots.txt and llms.txt live at the domain root. Your current price list, even if it is not published today. Three to five competitors you want to be measured against. And one person on your side who can approve wording, because otherwise the work sits in an approval queue.

If prices genuinely cannot be published, say so at the start. The work is still worth doing, but the strongest lever is gone and we build around it: service scope, typical project size, terms, coverage area, response time. Being specific about something is what makes you quotable.

On timing we say two things separately. Technical fixes take effect within weeks, because an assistant runs a fresh search for each answer and reads the page as it stands that day. Mentions and reviews take months, because they depend on other people writing about you, and that cannot be bought in a week. The fix work itself is about a week, and the audit is a single pass. If a supplier promises faster growth in mentions than that, ask how they measure it.

Start with the limit nobody can get around: a guaranteed place in an AI answer does not exist. Answers vary by question, models update, and the same question can return a different list next month. The honest promise is to measure how often you are named now, fix what raises the odds, and measure again. Anyone promising a guaranteed mention or a guaranteed position is selling air.

Four situations where we say no ourselves. One, there is no site yet, or it is being replaced in the coming months, in which case the money belongs in the new build. Two, your real problem is demand rather than findability. Technical readiness makes existing demand visible to a machine, it does not create demand. Three, you already know what is broken and only want it fixed, so skip the audit and order the work directly. Four, your score is already high and the remaining warnings are cosmetic.

Our own number is the fair example here. aiside.ee scores 95 out of 100 on the checker we publish, not 100. The price check warns on our own site because the euro sign is encoded by our rendering framework in a form the check does not recognise. A typical business site in our market scores under 50 on the same test. The distance between those two numbers is what this service is actually selling.