Jul 28,2026 
By Opttab 
Between January 2025 and May 2026, the share of US Google searches returning an AI Overview moved from roughly 15% to 43%. Over a comparable window, the share of ChatGPT answers carrying a web citation moved from about 1.3% to 6.8%.
Together those numbers are the defining tension of AI search in 2026: the answer layer is scaling far faster than the credit layer.
That gap is where most brands are losing.
TL;DR
- Google is not being replaced by AI. It is turning AI into the default interface of its own search product.
- Queries are getting longer and more conversational, so a keyword table now describes less of real demand.
- Citations are rising fast but remain uneven: 6.8% of US desktop ChatGPT answers, against 22.6% in travel.
- The page an AI cites is usually not the page a customer lands on: 65% of cited URLs sit two or three folders deep, while 58.8% of AI referrals arrive at homepages.
AI search is becoming part of Google’s default experience
Similarweb’s 2026 Generative AI Landscape, published 22 July 2026 and covered by TechCrunch five days later, tracks AI Overviews as a share of US searches. The series begins near 15% in January 2025 and reaches 43% in May 2026, its highest point.
The path was not a straight line. It peaks close to 40% in August 2025, drops to roughly 27% by October, then recovers. AI Mode traces a similar shape: visits rise from about 126 million in June 2025 to a peak near 320 million in March 2026, easing to 279 million by May 2026.
So the accurate claim is not that AI answers appear in most searches, but that they have become a default layer for a large and fast-growing minority of them, inside the product people already used.
A migration is something you can plan for. A quiet change to the interface your demand already flows through is not.
Search queries are becoming conversations
Similarweb reports the average US Google search grew about 5.4% longer after AI Mode arrived, from roughly 3.33 words to about 3.51. Over the same period, searches longer than five words moved from 13.6% to 16.1%, while one- and two-word searches declined.
Small numbers. Not small implications.
A short keyword string is ambiguous by design; the user resolves it afterwards by choosing which result to click. A longer, natural-language query carries that context inside itself: the constraint, the comparison, the use case, the budget.
That changes the unit of research. A keyword table describes what people type. A prompt universe describes what they are trying to decide.
Google is shifting from gateway to destination
For twenty years the results page was a departure lounge; its job was to hand you off. AI Overviews and AI Mode change that job: the page now answers, then invites you to continue in conversation. That is a real problem for publishers whose economics assume a click.
It is not the end of SEO. Every major AI answer surface depends on live retrieval, which depends on content being crawlable, indexed and credible. Organic visibility is now an input to AI visibility, not a competitor.
What has ended is the assumption that a ranking and a visit are the same event.
Citations are growing, but they remain uneven
Only 6.8% of analysed US desktop ChatGPT answers contained a citation in May 2026. That is more than five times the rate of a year earlier, and still a small share.
The category spread is far wider than the average suggests.
| Industry | Answers containing a citation |
|---|---|
| Travel | 22.6% |
| Retail | 13.5% |
| Sports | 10.7% |
| Finance | 8.0% |
| All analysed answers | 6.8% |
| Technology | 6.6% |
| Health | 6.3% |
| Media and entertainment | 6.0% |
| Education | 4.8% |
Similarweb’s source-mix data explains part of this: across the top 10,000 US desktop citations in May 2026, reviews and user-generated content supplied 28.9% of citations, and 54.1% inside travel conversations.
Opttab analysis, our interpretation rather than a Similarweb finding: the pattern fits categories whose answers depend on facts that move and can be checked: prices, availability, schedules, scores. Where an answer must be current, the model retrieves and shows its working. Where it is stable, it answers from memory.
The practical read: in a low-citation category, your visibility work is about being recommended, not linked.
The page cited by AI is often not the page users visit
This is the finding most likely to break your reporting.
Analysis by Aleyda Solis, published inside the Similarweb report, finds that around 65% of cited URLs sit two or three folders deep, with folder depth two alone accounting for 41.7% of citations. Meanwhile 58.8% of AI referral traffic lands on homepages.
The evidence and the entry point are different pages.
Your site now has three functionally distinct page types, each needing its own scorecard.
Evidence pages. Deep, specific, referenceable. What AI systems draw on to build an answer. Their job is to be accurate and quotable, not to convert.
Entry pages. Usually the homepage and top-level category pages. Where a user arrives after an AI has already introduced you. Their job is to confirm, within seconds, that the recommendation was correct.
Conversion pages. Where intent becomes pipeline or revenue.
Most analytics setups reward the entry page and never see the evidence page. Most content teams optimise the evidence page and are judged on the entry page’s numbers. Both measure something real. Neither measures the journey.
Zero-click does not mean zero influence
Similarweb puts the effect plainly: users were roughly 2.5 times more likely to visit a brand’s site within seven days of that brand being mentioned. In the report’s head-to-head brand pairs, the recommended brand drew two to four times the visits of the competitor it beat.
Almost none of that arrives labelled as AI traffic. It surfaces later as a branded search, a direct visit, a shortlist inclusion.
One caution, because this is where brands overclaim: that is an association measured across panels, not proof that a rise in your direct traffic was caused by AI. Test brand lift with holdouts rather than claiming it retroactively.
AI search is fragmented, not winner-take-all
Generative AI platforms averaged 9.5 billion monthly web visits worldwide between June 2025 and May 2026, up 70% year over year, with unique visitors up 57% to 655 million.
The more useful planning number is the overlap. Similarweb measured 494 million ChatGPT users worldwide between March and May 2026, of whom 95% also used Google. Gemini reached 310 million and Claude 86 million. Meanwhile ChatGPT’s share of standalone chatbot web visits fell from above 80% in late 2024 to just above 50% by May 2026.
Check how AI visibility tracking works across ChatGPT, Gemini and Perplexity
Almost nobody swapped one tool for another; people run several in parallel and choose by task. A visibility programme built on a single surface is measuring a sample of one.
What the data does not prove
- Panel data misses in-app assistants, API traffic and embedded AI features, so it understates total AI activity.
- Scope varies between charts: citation figures are US desktop, platform visits worldwide web, AI Overview share US. Quoting them as one dataset is a mistake.
- A mention preceding a visit does not prove it caused the visit.
- Measurement is contested. Adthena’s June 2026 data recorded AI Overviews in roughly 18% of US searches against Similarweb’s 43%, with no reconciliation published.
- Interfaces move monthly. ChatGPT’s 7 May 2026 link-format change reshaped referral behaviour inside a week.
What brands should do now
- Build a prompt and topic universe around the decisions customers are making, not only keyword volume.
- Track mentions, recommendations, citations, sentiment, prominence and competitor visibility separately. One combined score destroys the diagnosis.
- Map prompts to pages, so you know which page answers which question, and which questions have no page at all.
- Strengthen evidence pages with original data, direct answers, clean entity definitions and claims a machine can repeat safely.
- Make the homepage and top commercial pages ready for cold brand discovery. Many AI-sourced visitors arrive knowing your name and nothing else.
- Expand trusted third-party presence across publishers, review platforms, directories and communities. In several categories these supply most citations.
- Instrument the whole path: crawler activity, referrals, engagement, branded demand, conversions.
- Measure before and after every change, or you cannot separate your optimisation from the platform’s next update.
Get to know more about GEO / AEO and how they work.
That loop is what Opttab exists to close: tracking visibility across ChatGPT, Gemini, AI Overviews, AI Mode, Claude and Perplexity, and measuring whether the work moved citations, traffic and demand.
The takeaway
AI visibility cannot be read from rankings, citations or referral traffic alone. They are three layers of one journey, and a brand can win any while losing the others.
Measure them separately. Connect them deliberately.
The companies that win will not merely rank in search. They will become the brands AI systems understand, trust, recommend, cite and send users toward.
Opttab helps brands become the answer in AI search.
What percentage of Google searches trigger AI Overviews in 2026?
Similarweb’s 2026 Generative AI Landscape puts AI Overviews at 43% of US searches in May 2026, up from roughly 15% in January 2025. The series is volatile rather than a steady climb — it peaked near 40% in August 2025 and fell to about 27% in October before recovering. It is also worth knowing that other measurement firms report materially lower figures: Adthena recorded AI Overviews in roughly 18% of US searches in June 2026. No industry-standard measurement currently exists.
Is AI search replacing Google?
No. The data points the other way. Similarweb measured 494 million ChatGPT users worldwide between March and May 2026, and 95% of them also used Google. What is happening is that Google is absorbing AI into its own interface through AI Overviews and AI Mode, while users run several AI tools in parallel and choose between them by task. Discovery is fragmenting across surfaces rather than moving from one winner to another.
How often does ChatGPT cite sources?
About 6.8% of analysed US desktop ChatGPT answers contained a citation in May 2026, more than five times the rate of roughly a year earlier. Citation rates vary sharply by industry: travel reaches 22.6%, retail 13.5% and sports 10.7%, while education sits at 4.8%. Being cited and being recommended are different outcomes, and in low-citation categories a brand can influence answers frequently without ever appearing as a link.
How should brands measure AI search visibility?
Measure four layers separately, then connect them. First, mentions and recommendations — AI share of voice, recommendation rate, prominence and sentiment. Second, citations — how often you are used as evidence and which sources are cited instead of you. Third, website visits — AI referral traffic and which pages users actually land on. Fourth, commercial impact — engagement, branded demand, pipeline and revenue. Combining these into a single score hides the diagnosis, because the page that earns a citation is usually not the page that receives the visit.