Brand visibility in
AI search

1. Google isn't collapsing. The clicks leaving it are. Google's global search share rose from 89.34% in October 2024 to 91.27% in June 2026. Search volume is not what changed. The share of results that turn into a visit is.
2. Two thirds of searches now end without a click. In the US, the share of Google searches ending with no click was 60.45% in 2024 and measured 68.01% between January and April 2026.
3. The click effect of AI summaries is now causally established. In a randomized field experiment, outbound organic clicks fell 39.8% when an AI summary appeared. The same experiment also knocked down the defense that the lost clicks were "low quality."
4. The damage is not evenly spread. AI Overviews appear on 36% of informational queries, 8% of commercial ones and 5% of transactional ones. Sites that publish information and sites that sell products sit in different risk profiles.
5. AI traffic is small, but its quality flipped. Less than 1% of total traffic comes from AI interfaces. In Adobe's measurement, that traffic converted 38% worse than other channels in March 2025 and 42% better in March 2026.
6. Brand mentions are a stronger signal than backlinks. Across 75,000 brands, branded web mentions correlated at 0.664 while backlink counts stopped at 0.218. Correlation is not causation, but the ordering is clear.
7. No evidence for schema or llms.txt. In a controlled difference-in-differences experiment, adding JSON-LD did not increase AI citations. In a study of 137,210 domains, 97% of llms.txt files were never requested at all.
8. Google-Extended does not block AI Overviews. AI Overviews and AI Mode are fed by Googlebot. The only control is the nosnippet family, and it takes your regular search snippet with it.
9. Blocking bots without separating them costs you visibility. OpenAI's own documentation explicitly recommends allowing OAI-SearchBot, which feeds its search index. Governing training, search, and user bots with a single rule shuts off your chance of being cited in answers.
10. Turkey is a single-engine, narrow market. ChatGPT takes 94.49% of AI tool traffic there, the highest share in the world, but only 19.2% of individuals say they use generative AI. Yandex, at 10-15%, is another layer that almost no GEO tool sees.
11. Selling inside AI is not real yet. OpenAI deprioritized Instant Checkout in March 2026; the Google and Microsoft programs are US-only. For a brand in Turkey, the realistic goal today is not a sale. It is getting discovered and pulling qualified traffic to your own site.

The "Google is over" story going around the industry is not supported by 2026 data. Search volume is not what changed. The share of results that turn into a visit is.

Through 2024 and 2025, two claims competed in marketing: generative AI was ending search, or nothing was changing at all. The 2026 data falsifies both and points to something more interesting.

Start with what did not change. Google's global search market share fell to 89.34% in October 2024, its lowest level since 2015, then recovered to 91.27% in June 2026 Statcounter, June 2026. Desktop clickstream data points the same way: Google's US desktop share climbed from a Q4 2025 floor of 92.1% to 94.3% in Q1 2026 Datos, State of Search Q1 2026. Alphabet's search revenue grew 17% year over year in Q2 2026.

Now the part that did change. As of early 2026, OpenAI reported more than 900 million weekly active users for ChatGPT OpenAI, February 2026; the Gemini app reached 950 million monthly active users in July 2026 Alphabet Q2 2026. Pew Research finds 49% of US adults say they have used an AI chatbot, up from 33% two years ago Pew, June 2026.

So how much search does that usage amount to? The numbers diverge by a factor of 34, and that is not an error. It is a definition problem.

The honest version: limited to search-like queries, AI assistants amount to somewhere between 17% and 28% of classic search volume. Every figure below or above that range is an artifact of method.

The real break is in clicks, not queries

If search volume is not falling, why is brand traffic? The answer is zero-click. In the US, the share of Google searches that end without a click keeps climbing.

A second structural change landed in the same period. People ask AI assistants far longer questions than they type into search. In Semrush's 200 million user US panel, the length of prompts that trigger a search in ChatGPT went from 4.7 words to 8.7 words in a year Semrush, April 2026. In the same panel, queries in Google's AI Mode run about 1.8 times longer than classic Google queries.

That explains why keyword logic no longer covers it. An eight-word question does not show up in the volume tables of classic keyword research. What you need to measure is no longer a keyword. It is a prompt.

Four major studies give four different decline figures. The spread comes from method, and none of them is usable until you know what each one measured.

Through 2025, dozens of studies landed under the headline "AI Overviews are killing clicks." The figures ranged from 15% to 61%. In early 2026, one study settled the argument: a real randomized field experiment.

Agarwal and Sen randomly split US desktop Chrome users into three arms. In one arm a browser extension hid AI Overviews; in another they appeared as usual. The result: outbound organic clicks drop 39.8% when an AI summary appears, and the share of searches ending with no click rises 34.5% Agarwal & Sen, SSRN, April 2026. Most participants never noticed the extension was hiding anything.

What makes the study matter is that it also tested Google's standard defense. Google had argued that the lost clicks were "lower quality," visits the user would have bounced from anyway. In the experiment, visit-quality indicators were statistically indistinguishable across the two arms. The lost clicks were no different from the rest.

The damage is not evenly spread

This is the guide's most important strategic distinction. AI Overviews do not appear on every query, and where they do is systematic.

The practical read is clear: sites that publish information and sites that sell products are not in the same storm. AI summaries appear on informational queries more than four times as often as on commercial ones. A health portal and an ecommerce category page do not carry the same risk, and budgets and expectations should be split along that line from the start.

On the publisher side the picture is a clear loss. Chartbeat data drawn from more than 2,500 publisher sites shows global publisher traffic from Google organic down 33% in a year, and down 38% in the US Reuters Institute / Chartbeat, January 2026.

The number of visits arriving from AI interfaces is still small. What changed is what those visits do.

Start by sizing it correctly. In US panels, the traffic AI tools send to websites is less than 1% of total traffic SparkToro, June 2026. Conductor, measuring 13,770 enterprise domains, puts the average at 1.08% Conductor. Comparing 76,000 sites, Ahrefs finds Google sends 190 times more traffic than ChatGPT Ahrefs, February 2026.

That last figure has also been confirmed independently, by a completely different method. A peer-reviewed study in Marketing Science used first-party analytics from 973 ecommerce sites and found ChatGPT referral traffic roughly 200 times smaller than Google organic search Kaiser & Schulze, Marketing Science, April 2026.

The sign on quality flipped

Here is the finding worth pausing on. Adobe measures more than a trillion visits to US retail sites server-side and compares the conversion performance of AI-sourced traffic against other channels. The sign flipped within a year.

A second source points the same way: Shopify reported that in Q1 2026, orders from AI search grew roughly 13x year over year, that this traffic converted 49% better than organic search on product detail page sessions, and that average order value ran 14% higher Shopify, May 2026.

Usage is growing, exit clicks are not

Perhaps the most important structural finding of 2026: while visits to AI platforms grew toward 1.5 billion a month, the referrals those platforms send out stalled in a 240-280 million band between September 2025 and January 2026 Similarweb, June 2026. Referrals per visit are falling.

The first thing to break that plateau came on May 7, 2026: OpenAI made brand names clickable inside the answer text. In Similarweb's measurement, total ChatGPT referrals rose 157.7% within three weeks, and the share of visits landing on a homepage went from a 26-32% band to roughly 60% Similarweb, May 2026. That detail changes the character of the channel: ChatGPT stopped being a deep-link channel and became a brand discovery channel.

Every engine draws on a different universe of sources. Do not expect a single GEO strategy to work across all of them.

A generative answer is built from two layers of knowledge: what the model learned during training, and what it fetches from the web at answer time. The second is known as RAG, and it is where brands can actually intervene. Even if the model does not "know" you, you can still make the answer if you are findable and fetchable at the moment it is written.

So what do the engines fetch? On the same query, different engines cite radically different sources.

At the level of rank order, Reddit, YouTube, Wikipedia, and LinkedIn sit consistently near the top across all four major engines. On the size of the shares, though, no two studies agree. So rather than advise you to "be on Reddit," it is more honest to say measure the source profile of your own category.

How much of the answer comes from your site

This is why GEO work does not end on your own site. More than half of the answer comes from pages you do not control. Reviews, comparison sites, forums, marketplace product pages, and news coverage are all part of your brand story. That turns classic digital PR into a direct GEO lever.

Do Google rankings still matter?

Yes, but not on their own. Studies put the overlap between AI citations and Google's top 10 at somewhere between 38% and 76%:

The safe read: organic rank is still the single strongest signal, but roughly half of all citations come from outside the top 10. Ranking first on Google is neither a necessary nor a sufficient condition for AI visibility.

No single strategy works everywhere

Profound measured how much engines overlap in their sources across 100,000 different prompts: ChatGPT and Perplexity draw on the same source only 11% of the time. Cross-engine overlap runs between 6% and 16.4% Profound, July 2025. In SISTRIX's measurement across 82,619 prompts and 1.5 million snapshots, even Google's own two products, AI Overviews, and AI Mode, return different domains 83% of the time on the same query SISTRIX, May 2026.

Ghost citations

A joint study by Semrush and Kevin Indig measured an odd pattern: in 62% of AI citations the brand name never appears in the answer text, only the domain in the source list Semrush, June 2026. The sample is small (115 prompts, 3,981 domain appearances), so the pattern matters more than the exact figure. Platforms behave in opposite ways here: ChatGPT cites 87% of the time but mentions the brand only 20.7%, while Gemini cites 21.4% and mentions 83.7%.

The two most recommended tactics in the industry, schema and llms.txt, showed no effect in controlled tests. The evidence points somewhere else.

The founding academic text in this field is GEO: Generative Engine Optimization, presented at KDD 2024 by researchers from Princeton and IIT Delhi Aggarwal et al., KDD 2024. It tested nine different content interventions against a 10,000-query benchmark.

The paper's most striking finding is in the distribution: sites ranked fifth gained 115.1% from citing sources and 99.7% from adding quotations, while sites ranked first lost between 6% and 30%. The authors read this as "the potential to democratize the digital space." For small and mid-sized brands, that is where the leverage sits.

Brand signals beat link signals

Ahrefs measured Spearman correlations between AI visibility and a range of signals across 75,000 brands. The results partly invert classic SEO intuition.

The direction is still clear: how much your brand gets talked about on the web looks more decisive than how many links you have. In the same Ahrefs dataset, brands in the top quartile for web mentions average 169 AI Overviews mentions, while the next quartile lands at about a tenth of that. Roughly 26% of brands have no mentions at all.

Two popular tactics the tests found ineffective

Google's official position points the same way: "You don't need to create new machine-readable files, AI text files, or markup to appear in these features" Google Search Central, December 2025. John Mueller calls llms.txt "purely speculative for now" Search Engine Journal, June 2026.

Where the evidence points

The five items in this section are what technical teams misconfigure most often. Each one is verified against official documentation.

Know your bots: training, search, user

Treating AI bots as a single category is the most expensive mistake. They do three separate jobs, and governing all three with one rule costs you visibility directly.

Sources: OpenAI bot docs · Anthropic support · Perplexity docs · Google crawler list · Apple · Meta · Mistral · Amazon. The old anthropic-ai and claude-web tokens no longer appear in the official docs.

Blocking decisions: know what you are shutting off

Blocking AI bots wholesale looks tempting, and the training-data debate is a fair one. But a blocking decision made in a single line shuts off three different functions at once, and one of them is your visibility.

OpenAI states this plainly in its own documentation: OAI-SearchBot, the bot that feeds the search index, is one OpenAI recommends allowing, because it puts sites into ChatGPT's search features and is not used for training OpenAI bot docs. Perplexity and Anthropic draw the same distinction.

Plenty of crawling, few visits

Cloudflare tracks how many pages each platform crawls and how many visitors it sends back. The metric is extremely volatile, but the structural finding holds: classic search engines send a visitor every few crawls, while pure AI platforms send one visitor per hundreds or even thousands of crawls.

Where standardization stands

• The IETF AI Preferences working group is drafting a vocabulary that would extend robots.txt. The vocab draft reached its sixth revision in April 2026; the attach draft expired in May 2026. As of July 2026 no RFC has been published; the group is targeting August 2026 IETF aipref.
• Cloudflare's Content Signals Policy adds a Content-Signal: search=yes, ai-train=no line inside robots.txt, and as of September 2025 Cloudflare had added it by default to more than 3.8 million domains Cloudflare, September 2025. It is a statement of preference, not technical enforcement; no AI company has a verified commitment to honor it.
• Cloudflare's September 15, 2026 policy will block training-class and agent-class AI crawlers by default on ad-supported pages; the search class stays open by default. Scope: new customers, new sites from existing customers, and the free tier TechCrunch, July 2026. Bots that do both search and training (Googlebot, Applebot, Bingbot) are judged on their full behavior, and the most restrictive rule applies. Brands whose sites sit behind Cloudflare should put that date on the calendar.
• Monetization Gateway, announced in July 2026, is a separate product that aims to charge for API and tool calls as well as content; it is not generally available yet Cloudflare, July 2026. Pay Per Crawl, announced earlier, remains in beta.

AI answers are probabilistic. That makes some metrics meaningless, and others meaningful only with enough repetition.

This section may be the part of the guide that saves you the most money, because a large share of the reports on the market are statistically empty.

SparkToro and Gumshoe ran 12 prompts on three platforms 60 to 100 times each with 600 volunteers, 2,961 runs in total. The result: the odds that ChatGPT or Google's AI returns the same brand list in any two answers are under one percent. The odds of returning that same list in the same order are about one in a thousand SparkToro, January 2026.

The good news is that the visibility percentage is surprisingly stable with enough repetition. In the same study, after 60 to 100 runs, strong brands in narrow categories settle in a 55-97% band, and in broad categories in a 30-40% band.

How many measurements are enough

This is a statistics question, and the answer comes from the standard binomial margin of error formula: margin of error ≈ 0.98 / √n.

In practice the observation count is the product of four dimensions: how many prompts you track, how many phrasings you use for the same question, how many models, how many repetitions. For example, 10 prompts × 7 phrasings × 7 models gives 490 observations in one round and roughly ±5 points of precision.

How fast sources change

The second thing that sets your cadence is volatility: how fast the engines' source lists change. SISTRIX tracked 82,619 prompts weekly for 17 weeks.

The most constructive finding in the same study: 86.5% of prompts have a stable core of a few domains, while the remaining sources turn over about 89% week to week. So the goal is not to circle the edges but to get into the core. How long you stay in the core varies sharply by content type: in AI Mode, video and YouTube content stays 24% of the time, big tech sites 16%, Wikipedia 12%, and news media only 1.4%. News content is, in effect, a one-way ticket.

A suggested measurement protocol

1. Build the prompt set. Balance questions that name your brand with questions that do not. Measuring on branded prompts alone makes visibility look higher than it is. Group category, comparison, problem-led, and persona-based questions separately.
2. Vary the phrasing. Real user prompts share an average semantic similarity of 0.081, so people ask the same thing in very different ways. Measuring one template does not represent reality.
3. Record the baseline. Take the first measurement before you change anything. Every claim you make later rests on it.
4. Keep mentions and citations separate. ChatGPT and Gemini behave in opposite ways on these two; collapse them into one number and you lose the signal.
5. Keep the model and channel breakout. A brand that is strong in ChatGPT and weak in Gemini is completely ordinary.
6. Set measurement frequency by volatility. Weekly is balanced for most brands. Measure more often and you get noise; measure less often and you catch declines late.

2026 walked back the promise 2025 made about selling inside AI. For a brand in Turkey, the right target is not the sale. It is showing up correctly.

In September 2025, OpenAI announced Instant Checkout, which let people buy inside ChatGPT OpenAI. Five months later, in March 2026, the feature was deprioritized and purchasing moved to retailer apps inside ChatGPT Digital Commerce 360, March 2026. In OpenAI's own words, the first release "did not offer the flexibility we were aiming for"; basics like multi-item carts, promo codes, shipping commitments, and state tax were never covered Forbes, March 2026.

On Google's side, the Universal Commerce Protocol was announced in January 2026 and built with names like Shopify, Etsy, Target, and Walmart, but it is open in the US only, and only to selected merchants Google Merchant Center. Microsoft Copilot Checkout also opened in the US in January 2026 Search Engine Land, January 2026.

Product data: what you can actually do today

Checkout may have retreated, but the product feed is still what ChatGPT uses to discover and display products. OpenAI's feed spec is public, and a product missing a required field is dropped outright OpenAI feed specification.

Google Merchant Center's data quality rules run on similar logic, and most ecommerce brands in Turkey already have that plumbing in place. The one difference: the same data now works not just for ads, but for how AI describes your product.

Why the product page is the hardest page to read

The page where the purchase decision happens is, on most sites, the page AI reads worst. The reasons repeat: price, stock, variant, and delivery data load later via JavaScript; the product description is a copy of the manufacturer's text; the specs are baked into an image. All three leave the page readable for people and invisible to machines.

Fixing these three usually pays off faster than producing new content, because the page already exists and already gets traffic. The checklist is short:

• Are price, stock status, and variant data visible in server-rendered HTML?
• Are the specs text, or are they inside an image?
• Is the product description a copy of the manufacturer's text, or is it distinctive?
• Are shipping, return, and warranty terms available as readable text from the product page?
• Are reviews and ratings in the page source, or inside an embedded widget?

Turkey is one of the most single-engine AI search markets in the world. At the same time, adoption is still low and the supply of Turkish content is structurally thin.

Euronews Türkçe report, January 2026 (Digital 2026 compilation, October 2025 data) · TÜİK Household Information Technology Usage Survey 2025 · Statcounter Türkiye (May 2026) · W3Techs

The Yandex blind spot

Yandex's share is the most important data point separating Turkey from Western Europe. Statcounter puts it in a 10-15% band month to month, while Bing sits at a very low 1.33% Statcounter Türkiye. Two things follow:

• To the extent that the search layer behind Copilot and ChatGPT runs on the Bing index, being poorly indexed in Bing creates a quiet loss in Turkey. A low share in Bing does not make Bing unimportant.
• Almost no GEO measurement tool tracks Yandex. For a brand operating in Turkey, that means a tenth of the market sits outside measurement.

Turkish: the structural disadvantage and the opening

Turkish is an expensive language for language models. Measured on the 6,200-question TR-MMLU set, the same text took 434,526 tokens in one model's tokenizer and 561,866 in another, so model choice swings token cost in Turkish by 29% Bayram et al., arXiv:2502.07057. More striking: even the best tokenizer recognizes only 50.67% of Turkish words as whole units. The correlation between that rate and Turkish accuracy runs at 0.90, very high.

Bayram, Fincan, Gümüş, Karakaş, Diri, Yıldırım · arXiv:2502.07057 (February 2025, revised July 2025). Note: model size does not decide Turkish performance; Gemma-2 27B beats LLaMA-3.1 70B.

The flip side of this table is opportunity. Only 1.6% of the web is in Turkish, far below Turkey's population and economic weight. On Turkish queries, models struggle to find quality material to draw on. A brand publishing clean, current, verifiable Turkish content in its category can become a core source at a speed that would not be possible in English-language markets.

The size of Turkish ecommerce

Republic of Türkiye Ministry of Trade, ETBİS: Outlook for Ecommerce in Türkiye 2025 (May 12, 2026). TL-based growth rates carry inflation, so use the dollar series when comparing across years.

The legal frame: using a competitor's brand name

On the data protection side, the Personal Data Protection Authority published its Guide to Generative Artificial Intelligence and the Protection of Personal Data in November 2025 KVKK, November 24, 2025. The guide's "accuracy and currency" principle is directly relevant to GEO: incorrect personal data an AI produces about your brand can stop being a commercial visibility question and turn into a compliance one.

Order matters. Teams that act before they measure never learn what worked.

How this guide was built

• All data was verified against primary sources in July 2026. Secondhand write-ups were skipped wherever the original was reachable.
• Conflicting studies of the same thing were not collapsed into one number; a range is given, and the reason for the gap (denominator, panel, date) is stated.
• Vendor-sourced data is flagged. Companies that sell measurement tools have an incentive to prove that what they measure matters.
• Peer-reviewed academic sources and randomized experiments were weighted above observational vendor studies.
• Widely repeated claims that could not be verified were left out. Among them: the "schema lifts citations 44%" figure attributed to BrightEdge (no such study could be found), the claim that Anthropic and Perplexity support llms.txt (not confirmed in official sources), "AI traffic converts 23x better" (drawn from a single company's own site), and the widely quoted percentages on traffic loss at publishers that block AI bots (primary source could not be verified).

Known gaps

• There is no public study measuring AI's source profile on Turkish queries.
• There is no published causal experiment showing that raising brand mentions raises AI visibility; everything we have is correlation.
• No independently verified GEO case study with a named brand and transparent before-and-after measurement could be found. Every case in circulation is agency or tool marketing.
• Measurement tools change their parsing methods from time to time. Two figures from the same provider on different dates cannot be read as a trend if the method changed.

Main sources

For definitions of the technical terms in this guide, see the Brandmetric glossary.