ChatGPT searched the web in 14.4% of answers. What your AI visibility prompts are really measuring

Does ChatGPT search the web when a buyer asks about vendors like you?

Sometimes, and less often than most reports assume. Ahrefs, which sells AI visibility tools, tracked 66.6 million free, logged-out ChatGPT responses on GPT-5.6 Luna in August and September 2026, across all countries, and found that 14.4% triggered a web search.¹ The other responses came from what the model already held: its training data and whatever it had absorbed about your category before the question arrived.

That split explains the pattern in your report. An answer that cites your pricing page came from a run where ChatGPT searched, found the page and used it. An answer that names a competitor and no source came from memory, and nothing you published last week could have changed it. The two kinds of answer sit in the same prompt set and move for different reasons.

Prior work and a disclosure

A disclosure before the detail. Sigzen AI sells AI visibility audits and programmes at published prices, so we have an interest in you measuring this carefully.⁹ The method below works whichever tool or team you use, and most of it is a spreadsheet exercise.

Earlier estimates exist. Sistrix⁷ and Nectiv, reported by Digital Information World,⁸ each published their own measurements of how often ChatGPT searches, using different samples, tiers and methods. This post adds a way to use the newest and largest dataset to sort your own prompts, rather than another estimate.

The question in the title is a buyer's question, so the answer has to be usable in a report. "Does ChatGPT search?" becomes "which of our prompts does it search for, and what moves those answers?". The sections below take the Ahrefs data, the model change of 7 and 8 October 2026 and our own measurement method and turn them into that report.

What Ahrefs measured, and what it left out

Ahrefs measured whether a response triggered a search, how many fan-out queries each searching response ran, and which pages those searches read. The sample was free, logged-out ChatGPT on GPT-5.6 Luna in August and September 2026, all countries, 66.6 million responses.¹ When ChatGPT searched, it ran an average of about 1.6 fan-out queries per searching response, about 15 million fan-out queries in all.¹

Three limits matter for a brand. The sample is logged-out, so it says nothing about paid tiers, which run a different model and can carry memory of earlier chats. It covers one model over two months. And the 14.4% is a share of all responses, including small talk and coding help, not a share of buyer questions.¹

An adjacent dataset measures the toggle

Other datasets measure something adjacent. Semrush, which also sells AI visibility tools, read US clickstream data from its own panel. It reported that ChatGPT enabled web search on 34.5% of queries as of February 2026, down from 46% in late 2024.⁴ That counts sessions on a logged-in US panel where search was switched on, which is a different question from whether a logged-out model chose to search on its own.

The two figures are not in conflict; they describe two populations and two mechanisms. A logged-in user can turn search on by hand, and many do. A logged-out visitor gets whatever the default model decides. Our view: quote the Ahrefs rate for default, logged-out behaviour and the Semrush rate for the search toggle, and never average them into a single "ChatGPT searches X% of the time" line.

Free, logged-out ChatGPT, August–September 2026

Most answers never touch the web

14.4%of 66.6M responses triggered a web searchAhrefs, Aug–Sep 2026 ¹

~1.6fan-out queries per searching responseAhrefs, Aug–Sep 2026 ¹

84.4%of factual questions cited an official source, 20,000 US promptsAhrefs, Oct 2026 ¹

Query fan-out

When an AI engine decides to search, it rewrites the user's prompt into one or more search queries of its own, runs them and reads the results. Each rewritten query is a fan-out. The pages the engine reads are the ones that rank for the fan-out, not for the buyer's original words.

Retrieval

Answering with material fetched at the time of the question, as opposed to recalling what the model learned in training. A retrieval answer usually carries citations; a memory answer usually does not.

Why earlier estimates of the search rate disagree

They disagree because each one measured a different ChatGPT. Sistrix tracked, from its own daily review of responses, how often answers drew on live web search and how that changed over time.⁷ Nectiv's study, summarised by Digital Information World, ran a large set of prompts across several industries through its own detection tool and watched for the signs of a search.⁸ Different tiers, different months, different models and different definitions of "searched".

The lesson for a buyer is that a search rate is a property of one model, one tier and one period. None of these studies is wrong; they describe different animals. Ahrefs' figure comes from the largest sample we have seen and the one closest to what a logged-out visitor meets, which is why this post builds on it.¹

Our view: do not put any single search rate into a board slide as "how ChatGPT works". Put it in as "what the free tier did in August and September", with the date and the model name, and plan to replace it when the next study lands.

Which of your prompts can your own pages influence?

The prompts that make ChatGPT search, which are mostly factual ones. In Ahrefs' separate test of 20,000 US prompts, 84.4% of factual questions about pricing, product features or company details cited at least one official source.¹ For comparison questions the share fell to 68.3%.¹ A factual prompt sends the model to check, and when it checks, the company's own site is the page it trusts most.

Opinion and recommendation prompts behave differently. "Which payroll software should a growing firm in Australia choose?" is answered from memory far more often. When the model does search for it, the fan-out queries look for lists and reviews rather than for your site. Your pricing page cannot win a prompt the model never searches for.

Similarweb's data points the same way from the other end. In its US data, about 6.8% of ChatGPT answers carried a citation in May 2026, up from about 1.6% in June 2025, with travel near 23% and professional services under 4%.³ Categories full of checkable facts draw citations; categories built on general advice do not. Our note on that sector split covers what to track where links are rare.

Three prompt types, three levers

Factual prompts about your price, features, locations or certifications are won by your own pages being correct, current and easy to read. Comparison prompts are won partly by your pages and partly by the third-party comparisons the fan-out finds. Recommendation prompts are won slowly, through the mentions, reviews and lists the model absorbed in training, and they move when the model does.

Share of ChatGPT answers citing at least one official source, by question type

Factual: pricing, features, company details84.4%

Comparison questions68.3%

How do you split a prompt set by search tendency?

Label every prompt before you run it, then confirm the label from the runs. The label comes from the prompt's shape: a request for a fact, a comparison, a recommendation or a how-to. The confirmation comes from whether the answers carried citations across repeated runs. We use four steps.

Splitting a prompt set into memory and retrieval groups

1. Classify by shape

   Tag each prompt as factual, comparison, recommendation or how-to before any run.

2. Run on three days

   Each prompt on each engine, three separate days, so one lucky search cannot set the label.

3. Record citation presence

   Per run: did the answer carry any source link, and was one of them yours?

4. Assign the group

   Cited on every run: retrieval. Cited on none: memory. Anything between: mixed, reported on its own line.

Prompt wording matters more than it used to. Semrush compared January–February 2025 with January–February 2026 in its US clickstream data. Search-enabled ChatGPT prompts nearly doubled in average length, from 4.7 to 8.7 words, while non-search prompts nearly halved, from 24.9 to 13.5 words.⁴ People now type search-shaped questions when they want the web and conversational ones when they do not. Write the prompt set the way your buyers type and the search tendency will follow.

Our guide to building a prompt set from real buyer questions covers where the prompts come from: sales calls, tickets, Search Console and communities. The split above is what you do once you have them, and it should be refreshed whenever the set is.

What fan-out means for which pages get read

A searching answer rests on a handful of queries and the pages they return. Ahrefs counted about 1.6 fan-out queries per searching response.¹ Each fan-out is a search query ChatGPT writes for itself, and the pages it reads are the ones that rank for that rewritten query, not for the buyer's original words.

Where the fan-out queries go

The rewritten queries lean factual. In Ahrefs' data, "price" rose from 6.04% to 8.88% of fan-out queries between August and September 2026.¹ In the same data, 6.8% of fan-outs used a site: operator, and among the most common site: searches, 75.9% targeted brand sites.¹ When ChatGPT checks a fact, it often goes straight to the company's own domain and asks a search engine for a page there.

Once it reads, it reads widely. Semrush's 2026 AI Visibility Index, covering 126 million US prompts from January to April 2026, found ChatGPT cited an average of 15 sources per response, against 3 for Gemini.⁶ Across 11.84 billion citations on eight engines (16 April–16 July 2026), Profound found about 57% went to company-owned sites, and 47% on ChatGPT.⁵ Our post on owned versus third-party citations sets a target per engine.

Our view: for a retrieval prompt, the page to fix is the one that ranks for the fan-out query. That is usually a plain, well-titled page such as "Pricing" or "Integrations" rather than a long guide, and it has to carry the fact in text, not in an image or a tab. Our AI visibility service starts with a map of which of your pages those fan-outs reach, and which return nothing.

Why memory answers and retrieval answers move for different reasons

A memory answer changes when the model changes; a retrieval answer changes when the web changes. That is the whole reason to score them apart. If your retrieval score rises after a pricing page rewrite, the rewrite probably did it. If your memory score moves the week a new model ships, the model did it, and your work did not.

Retrieval scores also move when OpenAI changes what its search prefers. In Ahrefs' Brand Radar data on US ChatGPT citations from March to September 2026, PubMed climbed from 64th to 4th among the most-cited domains and the FDA entered at 8th.¹ Over the same months, citations of Reddit fell 88% and citations of Wikipedia fell 80%.¹ A health brand that saw its community-driven citations fall and official sources rise was watching a retrieval shift, not a reputation shift.

Our guide to measuring across a model update gives the before-and-after method for any engine change: freeze the prompts, compare equal windows, report with intervals. The split in this post feeds it. A model update should move the memory group more than the retrieval group, and seeing that pattern in your own data confirms the cause.

Two groups in one prompt set

Memory prompts

• Recommendations, opinions, "which is best"
• Answered without a search, rarely cited
• Move when the model or its training changes
• Lever: mentions, reviews, lists, time

Retrieval prompts

• Price, features, company facts, availability
• Trigger a search; usually cited
• Move when pages, listings and the index change
• Lever: your own pages, correct and current

Does this data describe the ChatGPT your buyers use today?

No. It describes the previous default. The Ahrefs sample is GPT-5.6 Luna, which OpenAI made the default model for Free and Go users in the week of 6 August 2026.² On 7 October 2026 OpenAI introduced GPT-6 in ChatGPT: GPT-6 Sol for Plus, Pro, Business and Enterprise, with GPT-6 Luna reaching Free and Go from 8 October.² The study appeared on 12 October, four days after the model it measured stopped being the free default.¹

That does not make the data useless. It is the cleanest published baseline for how a default model decides to search. The pattern it shows, factual prompts search while recommendation prompts recall, is a property of the product design as much as of one model. It does mean the 14.4% is a number to re-measure, not to memorise.¹

Paid tiers differ again. GPT-6 Sol serves the paid plans, and logged-in users carry chat memory that the logged-out sample never had.² Our view: every run in your prompt set should record which tier and model it hit, and a search-rate figure without a model name is not evidence.

Which model the search rate describes

1. 6 Aug 2026

   GPT-5.6 Luna becomes the default for Free and Go users that week.²

2. Aug–Sep 2026

   Ahrefs' tracking window: 66.6M free, logged-out responses, 14.4% with a search.¹

3. 7 Oct 2026

   GPT-6 Sol for Plus, Pro, Business and Enterprise.²

4. 8 Oct 2026

   GPT-6 Luna reaches Free and Go, replacing GPT-5.6 Luna.²

5. 12 Oct 2026

   Ahrefs publishes the study.¹

What should your AI visibility report show?

Two scores per engine, plus the share of prompts in each group. The retrieval score is the share of retrieval prompts where you were cited or named. The memory score is the share of memory prompts where you were named. The mixed group gets its own line. The levers then read off the page: pages and listings for the retrieval line, mentions and time for the memory line.

Repeat the runs and keep the engines apart

Repeat the runs. A single run per prompt per engine cannot tell a memory prompt from a retrieval one, because the model's choice to search varies between runs of the same prompt. Our audit runs 150 buyer prompts across ChatGPT, Gemini, Microsoft Copilot, Google AI Overviews and Google AI Mode, each on three separate days, and reports with confidence intervals.⁹

Keep the engine split. The search-rate finding is about ChatGPT, and the other four engines decide to retrieve on their own rules. A prompt that sits in the memory group on ChatGPT may sit in the retrieval group elsewhere. Label per engine, and resist a single blended score.

Date everything. The group a prompt belongs to is a finding about one model in one window, not a permanent property of the prompt. When a model changes, re-run the classification before you compare scores, or a prompt that moved from memory to retrieval will look like a page win you never earned.

Our view: the two-line report is the one change that stops teams claiming credit for a model update and blaming themselves for one. A page fix that lifts the retrieval line is yours to keep. A memory line that moves on 8 October belongs to OpenAI, and the next model will move it again.