AI Visibility & GEO
AI visibility and generative engine optimization
AI visibility is how often, and how accurately, AI assistants name your brand when someone asks a question you should own. It is measurable, it changes from week to week, and it is no longer downstream of your Google ranking.
For twenty years the job was to rank. A user typed a query, got ten links, and picked one. That machine still runs, but a growing share of questions now ends in a synthesized answer with three or four citations, and the rest of the web is invisible. If your brand is not in that answer, you were not considered. There is no page two to fall back on.
What is AI visibility?
AI visibility is your brand's presence inside generated answers from systems like ChatGPT, Google AI Overviews, Perplexity, Claude, Copilot and Gemini. It breaks into four things you can count separately.
- Mentions. The assistant names your brand, with or without a link.
- Citations. The assistant links to your domain as a source.
- Share of voice. How often you appear compared with the competitors that show up for the same question.
- Accuracy. What the assistant says about you. A confident wrong answer about your pricing or your services costs more than silence.
Most teams start by asking whether they are in there. That question has no useful answer because the same prompt returns different results on different runs, in different countries and in different languages. Visibility is a rate, not a state.
How is GEO different from SEO?
The unit of success changes. SEO optimizes a page for a query to win a position. GEO optimizes a passage for an intent to win a citation. You will see the same work sold as answer engine optimization, LLM SEO or AI search optimization. The differences are mostly marketing.
Three consequences matter in practice. One question becomes many, because AI systems split a question into sub-questions and assemble the answer from all of them, so you compete for a dozen adjacent questions you never targeted. What others say about you carries more weight, because models are grounded on the whole web and not on your site. And measurement becomes statistical: answers vary between runs, so you sample and track a trend.
Does ranking on page one still get you cited?
Less than it used to, and the drop is steep.
Ahrefs data published in 2026 found that pages ranking in the organic top 10 accounted for 38 percent of Google AI Overview citations, down from 76 percent in July 2025. The rest split roughly evenly between pages ranking 11 to 100 and pages that do not rank in the top 100 at all. BrightEdge, using a different methodology, reported an even lower overlap. The two numbers disagree, which is the point: nobody has a precise figure, but every dataset moves the same direction.
A strong ranking still helps, because a page Google trusts enough to rank is a page it can also retrieve. But it is now entirely normal for a page at position 40 to be cited while position 1 is ignored. That is good news if you are not the incumbent and uncomfortable news if you are.
Are AI answers costing you clicks?
Yes, and the effect is well documented.
Pew Research Center analyzed 68,879 Google searches from 900 US adults in March 2025 and found that when an AI summary appeared, users clicked a traditional search result on 8 percent of visits, compared with 15 percent when no AI summary was present. Clicks on links inside the summary itself happened on 1 percent of visits. Users also ended their session more often after an AI summary, 26 percent against 16 percent.
Ahrefs measured the same effect from the ranking side across 300,000 keywords: a 34.5 percent lower click-through rate for the top-ranking result when an AI Overview was present.
Your content is being read at industrial scale and the traffic is not coming back in proportion. The objective moves: from being the link someone clicks, to being the source the answer is built from.
Is traffic from AI assistants worth less than search traffic?
The visits are fewer, and the evidence does not suggest they are worse.
Semrush studied more than 500 high-value B2B topics in 2025 and reported AI search visitors converting at 4.4 times the rate of traditional organic visitors. Amsive ran a separate analysis and found no statistically significant overall difference, with a modest edge for LLM traffic on B2B sites. We quote both because they disagree, and because you should distrust any agency that quotes only the flattering one. Whether it holds for your business is an empirical question about your analytics, not about a benchmark.
What actually makes an AI system cite you?
Nobody outside the model providers knows the weighting, and anyone who tells you otherwise is selling. What is observable, repeatedly, across published research and our own testing, is a set of properties that cited pages tend to share.
Retrievability
The content has to be fetchable and readable without executing your JavaScript. Client-side rendered copy, content locked behind interactions, slow responses and overly broad robots.txt rules all remove you from consideration before quality is ever assessed.
Extractable structure
Question-shaped headings, an answer in the first sentence under each one, and self-contained passages. If a paragraph only makes sense after reading the three above it, it will not be quoted.
Specificity
Numbers, dates, named sources, defined terms, prices, methods. Generic claims are unquotable because they are interchangeable with every competitor's generic claim. This is the single biggest gap on most corporate sites.
Corroboration off your own domain
What comparison articles, industry directories, review platforms, forums and press say about you. A model that finds one claim on your site and nothing anywhere else has no reason to repeat it.
Entity clarity
Consistent brand naming, complete Organization structured data, correct sameAs references, matching details across every profile you own. Models need to know that the company in the article and the company on the website are the same company.
Maintenance
Visible publication and update dates, facts that stay current, and pages that get revised rather than left to rot. Stale figures get you cited once and then dropped.
What does not work
Keyword density. Mass-produced AI text, which reads exactly like the training data and adds nothing worth retrieving. Hidden instructions aimed at models, which are detected, ignored and a reputational liability if found. And buying a dashboard, which measures the problem without changing it.
Why the language of the question changes the answer
If you sell in Germany, Austria or Switzerland, measuring your AI visibility in English tells you close to nothing.
Ask an assistant the same commercial question in German and in English and you frequently get different brands, different sources and different confidence. The German answer is grounded in German-language sources, German industry media and German directories, most of which international brands have never touched.
This is where being based in Hamburg stops being a footnote. We run and read the German-language prompt sets, we know which German sources the assistants actually pull from, and we produce the German-language material that gets picked up. For an international marketing team that owns the DACH region from London, New York or Amsterdam, that is the part that cannot be handled remotely with an English-only tool.
How do you measure AI visibility?
We run two streams in parallel.
Prompt monitoring. We build a set of buying-intent questions your customers would actually type, in the languages of the markets you sell to. That set is run on a fixed schedule across several assistants. For every run we record whether your brand is mentioned, whether your domain is cited, which competitors appear instead, and whether the description of you is correct. Repeated sampling is not optional. A single check tells you nothing, because the same prompt on the same day can return a different answer.
Server-side reality. Log analysis for AI crawler activity, so we can see what is being fetched and what is being blocked, plus analytics segmentation for referrals from assistants. That is where the traffic and conversion question gets settled with your data rather than someone's benchmark.
We run the same prompt set on our own domain before we run it for a client, on a recurring schedule, which is also how we know what the reports look like when the numbers barely move. Our AI leaderboard is rebuilt from its data sources on a daily schedule, and each entry carries the date of its last update, so you can see how current the figure you are reading is.
How we work
- 01
Measure before you change anything
Prompt set, baseline over several runs, competitor set, crawler and referral audit. You get a starting number and the list of questions where competitors are cited instead of you. Without a baseline there is no way to tell later whether anything worked.
- 02
Fix retrieval
Rendering, speed, crawler access, structured data, internal linking, entity consistency. Unglamorous and usually the fastest measurable gain because a lot of sites are simply not being read properly.
- 03
Build passages worth quoting
We restructure existing pages into answer-first blocks with question headings, then add the specific, sourced material that is missing. Often that means publishing the numbers and methods your competitors keep vague.
- 04
Earn corroboration off your own site
Presence in the directories and comparison sources the assistants demonstrably pull from, plus the editorial work that gets your claims repeated somewhere other than your own homepage.
- 05
Re-measure and report
Same prompts, same schedule, tracked over time, with competitor movement alongside your own. Monthly reporting, quarterly review of the prompt set as your market and the models change.
What we do not promise
We do not promise citations, positions or a number. Model providers change retrieval behavior without notice, sometimes overnight. Google switched the default model behind AI Overviews in early 2026 and citation patterns shifted measurably within weeks. Anyone guaranteeing you a spot in an AI answer is guaranteeing something they do not control.
What we commit to is a defined prompt set, a documented baseline, a fixed cadence of measurement, and reporting that shows movement in both directions.
Where to start
With a baseline. Before anything is changed we measure where you stand today, in the languages of the markets you sell to, and produce the list of questions where competitors are cited instead of you. That measurement is the reference point everything later is compared against.
Tell us the markets you sell into and the questions your customers ask, and we will come back with a proposal for the prompt set and the measurement cadence. AI visibility is one part of our AI consulting and automation work, and it combines well with a website redesign when the technical foundation is the blocker. Our services are offered to businesses only and all prices are net of VAT.
Frequently asked questions
What is AI visibility?
AI visibility is your brand's presence inside generated answers from systems like ChatGPT, Google AI Overviews, Perplexity, Claude, Copilot and Gemini. It breaks into four things you can count separately: mentions, where the assistant names your brand; citations, where it links to your domain; share of voice against the competitors who appear for the same question; and accuracy, meaning what the assistant actually says about you. A confident wrong answer about your pricing or your services costs more than silence.
What is generative engine optimization (GEO)?
Generative engine optimization is the practice of making your content retrievable, quotable and corroborated enough that AI systems use it when they build an answer. You will see the same work sold as answer engine optimization (AEO), LLM SEO, LLMO or AI search optimization. The differences are mostly marketing. What it is not is a settings change, a plugin or a schema snippet you paste once. There is no ranking dial.
How is GEO different from SEO?
The unit of success changes. SEO optimizes a page for a query to win a position. GEO optimizes a passage for an intent to win a citation. Three practical consequences follow: one question becomes many, because AI systems split a question into sub-questions and assemble the answer from all of them; what others say about you carries more weight, because models are grounded on the whole web and not on your site; and measurement becomes statistical, because answers vary between runs, so you sample and track a trend rather than reading a single rank.
Does ranking on page one still get you cited?
Less than it used to. Ahrefs data published in 2026 found that pages ranking in the organic top 10 accounted for 38 percent of Google AI Overview citations, down from 76 percent in July 2025. The rest split roughly evenly between pages ranking 11 to 100 and pages that do not rank in the top 100 at all. A strong ranking still helps, because a page Google trusts enough to rank is a page it can also retrieve. But it is now entirely normal for a page at position 40 to be cited while position 1 is ignored.
Are AI answers costing us clicks?
The effect is documented. Pew Research Center analyzed 68,879 Google searches from 900 US adults in March 2025 and found that when an AI summary appeared, users clicked a traditional result on 8 percent of visits, compared with 15 percent when no AI summary was present. Clicks on links inside the summary itself happened on 1 percent of visits. Ahrefs measured a 34.5 percent lower click-through rate for the top-ranking result when an AI Overview was present, across 300,000 keywords.
Is traffic from AI assistants worth less than search traffic?
The visits are fewer, and the evidence does not suggest they are worse. Semrush studied more than 500 high-value B2B topics in 2025 and reported AI search visitors converting at 4.4 times the rate of traditional organic visitors. Amsive ran a separate analysis and found no statistically significant overall difference, with a modest edge for LLM traffic on B2B sites. We quote both because they disagree, and because you should distrust anyone who quotes only the flattering one. Whether it holds for your business is a question about your analytics, not about a benchmark.
What actually makes an AI system cite you?
Nobody outside the model providers knows the weighting. What is observable across published research and repeated testing is a set of properties cited pages tend to share: the content is retrievable without executing JavaScript, it is structured so passages survive being lifted out of context, it is specific enough to be worth quoting, it is corroborated somewhere other than your own domain, the brand entity is unambiguous, and the pages are maintained rather than left to rot.
Why does the language of the question change the answer?
If you sell in Germany, Austria or Switzerland, measuring your AI visibility in English tells you close to nothing. Ask an assistant the same commercial question in German and in English and you frequently get different brands, different sources and different confidence. The German answer is grounded in German-language sources, German industry media and German directories, most of which international brands have never touched.
How do you measure AI visibility?
Two streams in parallel. Prompt monitoring: a set of buying-intent questions your customers would actually type, in the languages of the markets you sell to, run on a fixed schedule across several assistants. For every run we record whether your brand is mentioned, whether your domain is cited, which competitors appear instead, and whether the description of you is correct. And server-side reality: log analysis for AI crawler activity plus analytics segmentation for referrals from assistants, so the traffic question gets settled with your data.
Can you guarantee we will be cited?
No, and neither can anyone else. Model providers change retrieval behavior without notice. Google switched the default model behind AI Overviews in early 2026 and citation patterns shifted measurably within weeks. Anyone guaranteeing you a spot in an AI answer is guaranteeing something they do not control. What we commit to is a defined prompt set, a documented baseline, a fixed cadence of measurement, and reporting that shows movement in both directions.
How long does it take before anything changes?
Retrieval fixes can show up within weeks, because they remove reasons you were never considered in the first place. Corroboration work off your own domain runs on the timeline of the publications and platforms involved, so months rather than weeks. Anyone promising a fast, guaranteed change to what a model says is describing something they cannot control.
Do we need this if our SEO is already good?
Good SEO is the foundation and it is not the whole job. A page that ranks is a page that can be retrieved, which helps. But the top 10 now account for a minority of AI Overview citations, and a well-ranking page written in vague corporate language is still unquotable. The work that remains is structural and editorial: making passages that stand on their own and saying things specific enough to be worth repeating.