The state of GEO in 2026, and who Baseline Labs are

AI answers stopped being a feature and became the default

The single clearest number: AI Overviews went from appearing in 15% of searches to 43% in twelve months, on Similarweb data reported in July 2026. Google's separate conversational product moved the same way. AI Mode visits rose from 126 million in June 2025 to 279 million by May 2026, and at I/O in May 2026 Google said AI Mode had surpassed a billion monthly users, with query volume more than doubling every quarter since launch.

The category around Google grew too. Generative-AI platforms averaged 9.5 billion monthly web visits, up 70% year over year, from 655 million unique visitors, up 57%, comparing June 2025 with May 2026. Google's standalone Gemini app went from 400 million monthly active users in May 2025 to more than 900 million by May 2026, per Google's own I/O 2026 keynote post.

Now the part most write-ups skip. Those prevalence figures do not agree. Semrush measured AI Overviews on 6.49% of queries in January 2025. BrightEdge put it at roughly 48% by February 2026, and Google's own figure for the same month was about 50%. Similarweb's 43% is dated five months later than both. Three credible measurements of the same phenomenon land seven points apart, and the later one is the lowest. The panels differ in query mix and geography, and in what each one counts as a search. The direction of travel is not in doubt. The level is, and anyone quoting a single AI Overview prevalence figure as settled fact is overstating what the data supports.

The map also closed. France was the last major Western holdout and only received AI Overviews and AI Mode on 22 July 2026, gated on publisher opt-out controls and neighbouring-rights payment commitments after €750 million in fines across 2021 and 2024 in the same dispute. There is no significant market left where you can treat AI answers as somebody else's problem.

How an answer gets assembled, and who gets named in it

The retrieval layer underneath these answers is now changing faster than the answers themselves look like they are changing. In January 2026 Gemini 3 became the default model behind AI Overviews globally, after testing from December 2025, and the "Show more" control started dropping users straight into AI Mode's chat interface. By I/O in May 2026, Google had gone further and folded the two products together, bringing AI Overviews and AI Mode into one AI Search experience with Gemini 3.5 Flash as the new global default model. Two default-model swaps and a product merge inside four months. Your citation rate can move without you touching a page.

Citation behaviour varies enormously by engine, which is the practical reason a single "AI visibility score" is close to meaningless. Semrush analysed 126 million US AI search prompts between January and April 2026 and found ChatGPT cites an average of 15 sources per response against Gemini's 3. A five-fold difference in how many slots exist. The same page can be a comfortable ChatGPT citation and completely absent from Gemini, and neither result tells you anything about the other.

ChatGPT also became much more willing to cite anything at all. Citation presence in its responses rose from about 1.6% in June 2025 to roughly 6.8% by May 2026, more than a fourfold increase. Its outbound behaviour shifted harder still: the share of ChatGPT desktop visits that landed on an actual outside webpage went from about 25% in March 2026 to nearly 60% by 30 May 2026, following a search update on 7 May. One vendor release changed how much traffic leaves ChatGPT more than a year of anyone's content work would have.

That last card matters more than its size suggests. Roughly 26% of ChatGPT responses now contain an ad. The answer surface is being monetised, on the same timeline as it is being adopted. Every assumption built on the idea that AI answers are a purely organic space has a shelf life.

Five things practitioners are still getting wrong

Treating Google rank as a proxy for AI citation. These are separate scoreboards now, and they diverge in both directions. A page can hold position one and never be named by ChatGPT or Perplexity, and pages with no meaningful Google visibility turn up as heavily cited AI sources. We wrote that up in detail: you can rank #1 on Google and be invisible to AI. If your reporting has one ranking column and no citation column, you are measuring half the problem.

Blocking AI crawlers wholesale. Training crawlers and answer crawlers are different populations with different user agents. A blanket Disallow aimed at "AI bots" removes you from ChatGPT and Perplexity search results, not just from training corpora. That is a decision worth making deliberately rather than by copy-pasted robots.txt: blocking all AI bots is actively costing visibility.

Using keyword volume as the unit of work. Engines fan a single question into sub-questions before retrieving anything, and most of those sub-questions have no search volume attached to them because nobody types them. A tool built on keyword volume is structurally blind to the queries that actually decide the answer. The better unit is the sub-question: stop optimising for keywords and optimise for sub-questions.

Betting the strategy on one engine. ChatGPT's share of generative-AI web traffic fell from roughly 76% in June 2025 to around 53% by May 2026 while the category itself grew. Gemini went from under 9% to around 27–28% in the same window, and Claude from barely 2% to close to 9%. ChatGPT losing 23 points of share while its absolute traffic rose is the kind of fact that breaks a single-engine plan quietly.

One more, and it cuts against our own product line. llms.txt is genuinely useful as cheap machine-readable exposure for assistants and agents. It is not a ranking lever, and Google has said its Search systems ignore it. We sell a generator for it and we still say that in public: llms.txt does nothing for Google Search, so here is what it is actually for. Structured data sits in a similar place. After Google deprecated FAQ rich results, schema's job shifted away from rich-result eligibility and toward making your entities legible to AI systems in general. The work still pays. The reason it pays changed.

Clicks down, referrals up, and the evidence is genuinely mixed

The strongest single piece of evidence on click impact is a randomised field experiment run across January and February 2026, two weeks per participant, around the France launch. When an AI Overview was shown, outbound organic clicks fell 38% and zero-click searches rose from 54% to 72%. Randomisation matters here. Most published AI Overview click studies are correlational panel comparisons, so they cannot separate the effect of the Overview from the kind of query that triggers one. This design can.

The other half of the ledger is growth. AI traffic to US retail sites rose 1,324% between October 2024 and May 2026, and travel 2,215% over the same nineteen months. Read those percentages carefully. They start from a base close to zero, so a four-figure increase is compatible with referral volumes that are still small next to organic search. Anyone selling you AI referral traffic as a replacement for organic on the strength of a 2,215% headline is using the base rate against you.

Which leaves the question of whether the visits you do get are worth more. Our honest answer is that the published evidence conflicts. Several studies report AI referrals converting at multiples of organic; others do not replicate it. We have written up both sides rather than the flattering one: AI sends less traffic, but the case that it is worth more has real holes in it. If a vendor gives you a single confident multiplier for AI conversion lift, ask which study it comes from and what the control group was.

The gap in the market is measurement, and that is what we build

Everything above is measurable in principle and almost nobody is measuring it. Semrush's 2026 survey found that 45% of marketing leaders cannot accurately measure their brand visibility inside AI-generated answers, and only 9% have the tools to track all the relevant metrics across platforms. That is not a taste gap or a budget gap. It is missing instrumentation, at the exact moment the surface it would measure became the default.

Baseline Labs builds tools for the generative-search era. We measure AI visibility, and we ship the structured data that lets AI engines understand and cite a site. The engines we track are ChatGPT, Gemini, Claude, Perplexity, Google AI Mode, DeepSeek and traditional Google organic, which is the set that the share figures above say you cannot safely ignore.

The measurement side. Visibility Scanner runs a saved list of real buyer questions against all seven engines and shows, per engine and per query, whether you are named, cited or absent, with trend lines across repeated runs of the same template. Brand Monitor sweeps 80+ web, social and news queries for mentions, scores sentiment on each, and surfaces the sources shaping how AI describes you alongside your mention share against named competitors. Active Competitors discovers your real rival set from the answers themselves by extracting co-mentions, rather than from a list you typed in, then benchmarks share of voice, sentiment and citation sources against it. Fan-out Audit maps the sub-queries an engine actually runs behind a seed question, using real fan-out data from Gemini's and ChatGPT's own APIs and proxy signals elsewhere. AI Traffic is first-party cookieless analytics that attributes visits to ChatGPT, Perplexity and Gemini referrals and ties them to on-site conversions. Rankings covers classic Google position, difficulty and CPC. Reputation Audit checks Google Business Profile, Trustpilot and TripAdvisor signals.

The fix side, because a diagnosis you cannot act on is a report. Schema Generator is free and needs no signup: paste a URL, get valid JSON-LD for 20+ schema.org types back in seconds. Rolling Schema re-crawls and regenerates that JSON-LD on a schedule so it does not go stale as the page changes, which is the part most schema tooling leaves to you. Schema Audit parses every JSON-LD, Open Graph and microdata block on a site and scores coverage and correctness. AI Vision Audit scores pages across seven AI-readiness signals, including bot access, structured data, entity authority and content clarity. Site Pulse grades reachability, metadata, content depth and rendering health, and catches pages that render blank to a crawler without JavaScript. llms.txt Generator builds a spec-compliant index file, with the caveat above attached.

Two things we would rather tell you ourselves. Our Backlinks graph is built entirely from our own crawler traffic across the other tools, not a licensed web-wide index, so the coverage is real but partial. And a benchmark number without an error bar is noise, so Active Competitors resamples paraphrase groups with a cluster bootstrap and only calls a share-of-voice move significant once it clears a 95% confidence interval. Competing tools in that category tend to show raw percentages with nothing around them. Given how much the underlying models moved between January and May 2026 alone, an unqualified percentage point of share is not a finding.

Where to start this quarter

Get a baseline before you change anything. Write down the twenty questions a real buyer asks before choosing in your category, run them across the engines your market actually uses, and record who gets named. That set of answers is your starting position, and without it every later movement is unattributable.

Then check you are reachable. An engine cannot cite a page it cannot fetch or cannot parse, and blank-rendering pages and accidental crawler blocks are still the most common reason a good page is invisible. Fix access before you touch content.

Then re-run the same questions on a schedule. A single scan tells you where you stand today. Repeated scans of the same template are what tell you whether a default-model swap in Mountain View just cost you a citation, and they are the only way to notice that before a revenue report does.

Run a mention scan on your brand

What people ask us about GEO