GEO and AI search · 10
GEO and AI Search Consulting.
GEO is SEO on new surfaces, and most claims about it come without a number. I measure what ChatGPT, Gemini and AI Overviews send and cite for your site, and what that is worth.
You hire me when
- Your board asks about ChatGPT.
- Nobody knows whether ChatGPT or Gemini mention you.
- A vendor sells you GEO tools or a GEO report.
- Your AI numbers moved and you do not know why.
What you get
- AI referrals per source, next to Google, from your own analytics.
- AI visibility check: your queries turned into prompts, mentions and citations counted, with the base printed beside every share.
- AI Overview presence per query type, repeated, with the date and the method.
- A claim ledger for every GEO tip you were sold: tested, plausible or unsupported.
- GEO KPIs, an owner per KPI, and a test backlog.
Why GEO needs a measurement before a plan
Behind most GEO talk sits one wish. My definition: “GEO is the desire by client to show up in ‘the AI’.” The wish is fair. The answers that come back for it are mostly guesses.
AI Overviews and AI Mode are built on Google’s own search. Google says so: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” (Google Search Central, generative AI optimization guide, last updated 2026-07-10)
ChatGPT is not Google, and its traffic is still small next to Google’s. It is also moving fast, as the data below shows. I told a team in a 2025 workshop: “These interfaces are not designed to send traffic to websites.” In 2026 the numbers on my panel stepped up within one week.
So a team needs counts before opinions: what do AI assistants send, what do they cite, and is the number stable? My goal: you know what AI sends and cites for your site, you trust the method, and you can tell a real change from a model update.
What goes wrong with GEO in 2026
These are the failures I measured in my GEO audits, workshops and re-runs from 2025 to 2026, and what I specify for each:
| What breaks | What I measured | What I specify |
|---|---|---|
| AI traffic planned on a guess | 2025, nine sites: AI sent 0.01 to about 1 percent of sessions | AI referrals per source, per site type, every month |
| A benchmark that is already old | 15 consumer-facing sites: 0.1 percent to 2.1 percent in 20 months; +45 percent in one week | Re-measure monthly, date the events, name the cause as proven or inferred |
| AI growth read as Google’s replacement | AI +85k sessions, Google organic -325k in a year, same 15 sites | Both lines on the same sites, side by side |
| A share on a base nobody checked | My own first report said about a fifth of 100 queries; the real base was about 50 | Rows, errors and base printed beside every share |
| AI Overviews: sometimes there, sometimes not | About 4 in 10 of 50 queries; the same query 15 minutes apart differed | Per query type, repeated, dated, with the method named |
| Visibility that is only your brand name | All brand-name queries visible; about 1 percent of the other 220 | Brand and non-brand as two numbers |
| A trend that is the model changing | Same 500 prompts: Gemini 1 in 10 to over 4 in 10 | Pinned model id, a control run, identical queries only |
| Vendor tips without evidence | About a dozen claims in one vendor article, two plausible, none with a number | A claim ledger: claim, source, test, verdict |
AI traffic planned on a guess
For my JETZT GEO/SEO keynote in January 2026 I pulled a full year of analytics for nine sites. AI assistants sent between 0.01 and about 1 percent of sessions.
- News sites were at the bottom, under 0.03 percent.
- The top of the range was about 1 percent.
- ChatGPT was 80 to 94 percent of the AI sessions.
Your board’s “everybody uses AI” is not a traffic number. Your own analytics is.
What I specify: AI sessions per source (chatgpt.com, perplexity.ai, gemini.google.com and others), per site type, from your own analytics, next to Google’s clicks. Then decide what a project is worth.
A benchmark that is already old
In October 2026 I re-ran the numbers on a wider panel, chosen by a written rule: 15 consumer-facing sites, no news. The AI share of non-direct sessions was about 0.1 percent in January 2025 and about 0.5 percent from September 2025 to April 2026. In September 2026 it was about 2.1 percent.
- AI sessions grew from about 8k to about 110k a month, roughly 14 times.
- ChatGPT was about 95 percent of them, Gemini under 4 percent.
- In the week of 11 May 2026 ChatGPT visits per 1,000 Google organic visits rose about 45 percent in one week.
- Engaged-session rate of AI visits stayed at 77 to 79 percent, so this is not bot noise.
The timing fits a ChatGPT change on 7 May 2026. Similarweb reported: “Since May 7th, ChatGPT began surfacing more prominent links to brands in its answers.” (Search Engine Roundtable, reporting Similarweb) That the change caused my step is my inference, not proof.
Paid clicks did not cause it: Fewer than 20 of about 740k AI sessions carried a paid medium.
Three news sites in the same pull went from 0.01 to 0.04 percent. Five other sites stayed at about 500 AI sessions a month, flat since January 2025. The same ruler gives opposite pictures by site type.
What I specify: one monthly AI line per site type, dated events next to it, and a label on every cause: proven, or inferred.
AI growth read as Google’s replacement
On the 15 consumer-facing sites AI sent about 85k more sessions over one year, from September 2025 to September 2026. Google organic sent about 325k fewer sessions on the same sites.
In January 2025 Google organic sent about 290 visits for every AI visit. In September 2026 it sent about 13. AI is closing the gap, and it is not replacing what Google lost.
I do not claim AI caused the Google loss, or that Google caused the AI gain. I claim the two lines must sit in one chart before anyone writes a business case.
What I specify: AI referrals and Google organic sessions on the same pages and sites, weekly, with the ratio. Absolute numbers beside every percent.
A share on a base nobody checked
My own first report on an AI Overview check said: about a fifth of 100 queries showed an AI Overview. When I re-ran the check in October 2026, I found that half the rows in the input file were empty. The real base was about 50 queries, and the December share was about a third.
A tool that reports a percent on the wrong base looks exactly like a tool that reports it on the right one. The re-run uses the 50 real queries, and my notes carry the corrected base.
What I specify: every share carries its base, its failed rows and its date. A re-run that cannot reproduce the base is not a result.
AI Overviews: sometimes there, sometimes not
In the October 2026 re-run, About 4 in 10 of 50 queries showed an AI Overview on mobile. In December it was about a third.
- All queries that had an overview in December still had one. A few more gained one. None lost one.
- The more than half of queries that showed none in either run were short and transactional.
- The site was cited as a source in about a third of overviews in December and about 1 in 10 in October. The two checks used different methods, a browser scrape and a search-results API, so I read that as a direction, not a drop.
In a 2025 session I showed a second effect: the same query, on the same device, 15 minutes apart, sometimes showed an overview and sometimes did not.
Google describes the pattern: “AI Overviews are only shown when our systems determine that it is additive to classic Search, and as such, often don’t trigger.” (Google Search Central, AI features and your website, last updated 2025-12-10) One check is a snapshot, not a rate.
What I specify: AI Overview presence per query type, repeated runs on different days, and Google’s own report next to it (see the questions below).
Visibility that is only your brand name
In December 2025 I checked 300 prompts built from one brand’s top Search Console queries, on two models through their APIs. The brand showed in about a quarter of queries. In July 2026 it was slightly less.
- The drop is noise. On the queries that were in both runs (a bit over half), the rate was about 27 percent both times.
- The visible queries were almost all queries that named the brand or a product. Every one of the about 80 such queries was visible. Of the other 220, about 1 percent were.
- The answers named any website at all in only about 1 to 3 percent of cases.
That is a brand-awareness check, not an authority check. Almost no query without the brand name brought the brand back.
What I specify: brand and non-brand queries as two numbers, the same queries in every run, and the question “who is cited when nobody typed our name”.
A trend that is the model changing
I re-ran the same 500 saved prompts for one site ten months after the first run. Gemini named the brand in about 1 in 10 answers in December and in more than 4 in 10 in October. A pinned OpenAI model stayed at about three quarters both times.
The brand had not changed. Google had moved the model behind the “latest” alias in January 2026 (Gemini API changelog). My re-run used the closest model I could afford. It named websites in almost half of its answers, against about 1 in 10 before. A control run with reasoning switched off sat in between.
Neither run used live search. Both measure what a model says from what it learned, not what the apps show. A GEO dashboard that does not pin the model id will show you this jump as your own success.
What I specify: model id and date in every row, a control run when the id changes, and trend lines only on identical prompts.
Vendor tips without evidence
For one team I took apart a search-engine vendor’s October 2025 article on how to get into AI answers. I extracted about a dozen claims and looked for a primary source for each.
- Two claims held up as plausible: schema markup, and classic SEO as a baseline.
- None was confirmed by an official document.
- None of the claims gave a number.
Google says it plainly for its own systems: “No third-party tool has access to our internal ranking or AI systems.” (Google Search Central, generative AI optimization guide, last updated 2026-07-10)
What I specify: every GEO tip goes into a ledger: the claim, where it comes from, what test would show it, and the verdict. Tips that cannot be tested stay out of the backlog.
How I work
- Count first. Before any workshop I pull AI referrals per source from your analytics and Google’s clicks for the same pages. In one 2026 pull the same ruler gave 0.04 percent for news and about 2 percent for consumer-facing sites.
- Date the events. I keep a log of what changed at the assistants and at Google next to your lines. In one pull a ChatGPT change on 7 May explained a step I would otherwise have credited to a project.
- Check visibility with a method you can read. Your top Search Console queries become prompts, written without knowing your brand, and I count brand mentions, links and source-list entries. The base, failed rows, model ids and dates are printed with every share. The tools are public on GitHub.
- Repeat, then compare. Every check runs again on identical prompts. I only call a change real when the identical queries show it.
- Test the claims you were sold. Each vendor tip goes through the claim ledger. I check Google’s own documentation first, and say “not documented” when it does not say.
- A GEO day with your team. The day runs on your numbers, not on slides about other sites. In a full GEO day in 2025 I walked one team through 30 days of its own AI traffic, source by source.
- KPIs with owners. You leave with KPIs, an owner for each, a test backlog and a monthly re-measurement date.
My own tools, checked the same way
I build the checkers I use, and I publish them: the AI Relevancy Checker and the AI Visibility Tracker are on GitHub.
The October 2026 re-run is where my own error showed up. I found the half-empty input file behind my first AI Overview number. I found that a jump in Gemini mentions was the model changing, not the site. I wrote both down, and I use the corrected numbers.
I also run open GEO days in Vienna. Both 2026 dates had 12 seats, and both were full.
Questions people ask me about GEO
Do we need an llms.txt file or special markup to get into AI answers?
Not for Google. Google: “Doing so will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.” (source) That sentence is about LLMS.txt files. On the AI features page: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” (source)
Does Google report AI Overviews and AI Mode in Search Console?
Yes, since 2026. Google: “As of August 31, 2026, we’ve rolled out these insights to all websites worldwide.” (source) It counts impressions, not clicks: “Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search.” (source) I read it next to your clicks and your analytics.
Which pages can show up in AI Overviews at all?
Google: “a page must be indexed and eligible to be shown in Google Search with a snippet”. (source) That is why the first check is classic technical SEO, see the technical SEO page.
Do we need to cut our content into small chunks for AI?
Google: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” (source) That is about Google. For other assistants I test it on your pages instead of repeating a tip.
How much traffic do we get from AI?
Measure it in your own analytics. In my data it was under 0.05 percent of non-direct sessions for news sites and about 2 percent for consumer-facing sites in September 2026. Your site type and your month decide. The method is on the search data page.
Can you get us into ChatGPT answers?
I cannot promise a place in any answer. What I can do is count when you appear today, for which kind of query, and whether that is stable. Google says no tool can see its ranking systems from outside. I treat claims about the assistants the same way: tested, or not believed.
Is GEO a new discipline, or SEO?
On Google’s side: still SEO. On the other assistants: too early to say. In my words in a 2025 workshop: “GEO is not yet a mature discipline. And it is not even sure if it will become one.” What is new is the measurement work, and that is what I do.
Sources for the Google quotes on this page, read on 2026-11-12: AI features and your website (last updated 2025-12-10), Optimizing your website for generative AI features on Google Search (last updated 2026-07-10) and the Generative AI performance report, Google Search Central and Search Console Help. The ChatGPT link change comes from Search Engine Roundtable, reporting Similarweb. The client numbers come from my own audits, workshops and re-runs in 2025 and 2026.
From real jobs
- 2026: AI assistants sent about 2 percent of non-direct sessions on 15 consumer-facing sites, up from 0.1 percent in January 2025. Three news sites: 0.04 percent.
- 2026: the same 500 prompts re-run after 10 months. Gemini named the brand in about 1 in 10 answers in December and in more than 4 in 10 in October. The model had changed, not the site.
- 2025 to 2026: two 300-query AI visibility checks for one brand. About a quarter of queries showed it, almost all of them queries that named the brand.
- 2025: a fact check of a search-engine vendor's GEO article. About a dozen claims, two plausible, none with a number.
- 2025: a GEO session for one team, and a full GEO day for another.
- January 2026: keynote at JETZT GEO/SEO, on nine sites with a year of analytics: AI sent 0.01 to about 1 percent of sessions.
Teams like yours
Some of our customersNews publishers
Dailies, agencies, broadcasters: Google News, Discover, training.
ExamplesMarketplaces and job platforms
Real estate, jobs, cars: list pages decide traffic.
ExamplesHealth and pharma
Health portals, medical education, pharma: YMYL topics.
ExamplesPublic sector and institutions
Government, chambers, the post: many stakeholders.
ExamplesWorkshops for this
All workshop formatsSomething else in the way?
How consulting works- External SEO lead
- Relaunch and migration SEO
- Technical SEO
- International SEO
- Web performance
- SEO for newsrooms
- Content strategy and search demand
- Search data and traffic forensics
- SEO agency cleanup
- Vibe coding training
- AI automation consulting
- AI-first transformation
- Custom tools for one job
- Design thinking
- Growth
- Growth strategy analysis
- SEO workshops and training
- Keynotes and talks
- Sparring and second opinion
- Dev, product and marketing mediator