87 AI Search Volatility Statistics

Ask Google AI Mode the same software-buying question ten days apart and it names about half the same brands: 48% of them, in the B2B software category we measure every week. ChatGPT names 58% of them again.¹ Across the whole category, those same ten days moved the median brand’s share of answers by less than three points.

Most teams read an AI answer the way they read a ranking, as a position that moves now and then and can be checked with a screenshot. An answer is closer to a draw from a pool of candidates the engine refills every time it is asked, so it moves a lot from one answer to the next and far less across a whole category.

What follows are the numbers on how much AI answers move, why they move, and how fast buyers are moving to them: our own weekly panel across two software categories, research from Qvery, our sister company, and the published studies, each read at its origin.

Highlights

Fewer than 1 in 100 repeat runs of a prompt on ChatGPT or Google’s AI return the same list of brands, and about 1 in 1,000 return it in the same order.²

48% of the brands Google AI Mode named for a software question were named again ten days later, and 58% on ChatGPT.¹

47% after three weeks, against 48% after four days: most of the change in an AI Mode answer comes from asking again, and very little from time passing.¹

56% and 74% of cited domains are new each week on Google AI Mode and ChatGPT,³ and in our category 8% of AI Mode’s cited pages survived ten days.¹

1 point was the median four-day move in a brand’s share of answers across a whole product line, and only 8% to 9% of moves beat the noise, against 5% by chance alone.¹

One reading in ten lands 11.8 points or more from the full figure when visibility is measured on 25 questions.¹

37.3% to 48.7% of ChatGPT runs came back with sources in Qvery’s June studies, and 93.83% to 98.88% two months later.⁴

51% of B2B software buyers begin their research with a chatbot more often than with Google, against 29% eleven months earlier.⁵

Any single AI answer is a draw, so one answer, or a small set of them, cannot tell a team where it stands. The picture the draws add up to is steadier. It still moves when a model or a product changes, and more of the buying is moving into the engines doing the drawing.

Ask Again and a Different Answer Comes Back

Our panel asked ChatGPT and Google AI Mode the same 9,000 unbranded software-buying questions on 13 and 23 September 2026, in one B2B software category, and Claude 900 of them. Matching each question to itself ten days later:

58%, 48% and 69% of the brands named for a question were named again, on ChatGPT, AI Mode and Claude.¹

51%, 36% and 64% of questions kept the same first-named brand.¹

7%, 4% and 11% came back with the identical list.¹

19%, 20% and 3% shared no brand at all with the first answer.¹

Google AI Mode moved most. Claude moved least, and it is the one engine we ask through its API rather than through the page a buyer sees, which says as much about the route as about the model. A consumer finance software category we also measure kept the same order over ten days: ChatGPT named again 66% of the brands, AI Mode 55%.⁶

Each question matched to itself across two weekly waves. Tracked vendors only, the partner behind the panel excluded.

Every study that re-ran the same prompt found the same pattern at its own speed:

Fewer than 1 in 100 repeat runs on ChatGPT or Google’s AI returned the same list of brands, across 2,961 runs by 600 volunteers.²

0.3% of day pairs on ChatGPT returned an identical set of brands for a prompt checked daily for a month, and 1.1% on AI Overviews; 41% of ChatGPT’s prompts never repeated a list once.⁷

30% of brands stayed visible from one answer to the next, and 20% across five consecutive runs.⁸

44 brands across 100 runs of one B2B software prompt on ChatGPT, about 10 per answer.⁹

Read as a sample, the list is doing what a sample does. An engine holds a pool of candidates for a question, larger than any one answer shows, and each answer draws a handful from it. Ask a hundred times and you meet the pool. Ask once and you meet whatever came out this time.

Empact Partners runs this panel for the partners in its Generative Engine Optimization workstream, GEO for short: the work of becoming a brand AI engines name when a buyer asks their category question. The same unbranded questions go to three engines every week and every answer is kept, so a question can be matched to itself.

Leon Claassen

Senior GTM Consultant @ Empact Partners

Asked the same software question ten days apart, Google AI Mode drops about half the brands it named and ChatGPT about four in ten. A screenshot of one answer, in a board deck or on a sales call, shows one draw from that pool and says nothing about your odds in it. The number worth reporting is how often you are drawn across hundreds of questions. Your position inside a single answer is not a metric.

Most of the Movement Is the Engine Drawing Again

If the churn were the market moving, it would deepen with time: a question asked three weeks later would have lost more of its brands than one asked four days later. In our category it barely does. From the same 9,000 questions asked on 2 September:

Google AI Mode named again 48% of a question’s brands four days later and 47% three weeks later.¹

ChatGPT named again 64% after four days and 57% after three weeks.¹

The first-named brand held on 37% and 35% of AI Mode’s questions over the same two gaps, and on 57% and 55% of ChatGPT’s.¹

A month out, ChatGPT named again 53% of the brands it had named for a question on 22 August.¹

So most of what changes between two answers is there on the first re-ask. On AI Mode, three weeks added almost nothing to it. On ChatGPT, time added a slow slide on top of the draw, from 64% to 57% over seventeen days.

Four days and three weeks after 2 September, one B2B software category. If the churn were the market moving, the right-hand dots would sit far lower than the left.

Qvery, our sister company, sees the same thing in its own data. It tracks millions of citations a month across ChatGPT and Google AI Mode. It is ours, so give its findings less weight than an outsider’s, and open the studies linked here to check the working yourself.

In its share-of-voice test, ChatGPT answered the same shopping question again and again inside one collection window, and two thirds of the cited sources moved between any two runs.¹⁰

The draw comes from how an answer gets built. Three steps sit between the question and the answer, and each can land somewhere else on the next ask:

Rewriting. ChatGPT turns a question into one or more search queries and may pass your general location to the search providers.¹¹
Fan-out. Google’s AI Overviews and AI Mode issue multiple related searches across subtopics for one response, and the two may use different models, so the links they show vary.¹²
Sampling. The model then writes the answer word by word from probabilities. Even with that randomness switched off, one lab got 80 different completions from 1,000 runs of one prompt, because the server batches each request with other people’s and the batch changes the arithmetic.¹³

Place and person change the pool

The same question from a different place, or a different person, draws from a different pool:

35% of domains repeated between runs of the same local query on AI Mode in the same city, and 24% across cities.¹⁴

9.2% of pages were shared by three runs of the same 10,000 keywords on AI Mode on a single day.¹⁵

28% of combinations of brand, persona and engine moved five points or more when a persona was added to a software prompt.¹⁶

A tracking setup that asks from one location, logged out, with no persona, measures one corner of that pool. That is fine as long as nobody reads the corner as the whole.

Asking an engine the same question twice samples it twice. Visibility is how often a brand comes up across hundreds of those samples, and one answer cannot show it.

The Pages Behind an Answer Turn Over Faster Than the Brands in It

Brands are the steady layer. The pages an engine cites underneath them rotate much faster, and over the same ten days in our category:

32%, 18% and 68% of the domains cited for a question were cited again, on ChatGPT, AI Mode and Claude.¹

19%, 8% and 64% of the exact pages were.¹

Week on week, 64% and 77% of the domains ChatGPT and AI Mode cited for a question were new.¹

On the same question on the same day, ChatGPT and AI Mode shared between 7% and 9% of their cited domains,¹ and named the same brand first between 31% and 34% of the time.¹

The deeper the layer, the less of it survives: on both engines that search the live web, the brand list keeps about half, the domains a third or less, and the exact pages a fifth or less.

Horizontal bar chart of what survived ten days on the same software question: brands named again 58% on ChatGPT, 48% on Google AI Mode and 69% on Claude; cited domains 32%, 18% and 68%; cited pages 19%, 8% and 64%. — One B2B software category. The two engines that search the live web as a buyer sees them lose most of their sources in ten days while keeping about half their brands.

Every published measurement of the source layer finds the same speed, whatever its ruler:

56% and 74% of the domains cited in Google AI Mode and ChatGPT responses are new each week, 54% to 59% in each of six countries, with no sign of settling over 17 weeks.³

40% to 60% of cited domains differ a month later for the same prompt, and 70% to 90% from January to July.¹⁷

70% of the time an AI Overview’s content changes between two observations, and 45.5% of its cited pages are new each refresh.¹⁸

6.7% of desktop pages cited in AI Overviews in the first week of October 2024 were still cited in the last.¹⁹

70% of the pages in AI Overviews could be expected to change within two to three months, in the first published read in February 2025.²⁰

About a quarter of yesterday’s sources were cited again by ChatGPT and AI Mode, against roughly three quarters by Perplexity.²¹

Google’s own two surfaces barely agree either. AI Mode and AI Overviews shared 13.7% of cited pages for the same query while reaching answers 86% alike in meaning.²²

A brand outlives the pages because several of them say its name, which is the logic behind the formula our GEO work runs on, GEO = UGC + Mentions: what real users say about a brand in public, plus where independent pages name it.

A brand named across dozens of the pages an engine rotates through keeps turning up whichever ones it draws this week. A brand that depends on one ranking page drops out the week that page rotates away.

Leon Claassen

Senior GTM Consultant @ Empact Partners

Over ten days in our category, ChatGPT kept a third of the domains it cited for a question and Google AI Mode under a fifth, while both kept about half their brands or more. The brand outlives the page because many pages carry it. So the durable work is getting named on many of the pages an engine cycles through. The page you rank for today is a draw you won this week.

Across a Whole Category, the Picture Holds for About a Week

Step back from the single question to the whole category, which is where a marketing team’s number lives, and most of the churn cancels out.

The median brand barely moves

Across nine product lines of about 1,000 answers each, between 2 and 6 September:

A brand’s share of answers moved a median of 1 point.¹

8% and 9% of moves on ChatGPT and AI Mode were bigger than sampling noise alone would produce, where chance puts about 5% past that line.¹

SparkToro found the same split inside a single prompt. ChatGPT named one West Coast cancer hospital in 69 of 71 answers and put it first in 25.² Its visibility held while its place in the list changed from answer to answer.

Brands at 5% visibility or more in either wave, one B2B software category. A move counts as beyond noise past 1.96 standard errors of the difference.

Leaders hold, and the tail rotates

75% of the time a brand in its product line’s top five was named again four days later on ChatGPT, and 61% on AI Mode; every other brand, 49% and 32%.¹

21% and 10% of the brands a question drew over four weekly asks were named every week, on ChatGPT and AI Mode, while 41% and 52% appeared once.¹

15% of AI Mode’s questions kept the same first-named brand all four weeks, and another 15% named a different one each week.¹

86.5% of AI Mode prompts keep a stable core of one to five cited domains, while 89% of the rest rotate weekly.³

ChatGPT’s core brands for a prompt changed by 13% from one day to the next, against 78% for its tail brands.⁷

Qvery’s same-day shortlist test found the same shape in two SaaS categories: across five runs of the same questions, 4 brands in project management and 2 in CRM made every run’s top five, and some brands made one run’s top five and no other.²³

A category has a core the engine keeps returning to and a carousel it spins around it. A leader’s visibility is sticky because it sits in the core. The fight for most brands is in the carousel, which is exactly the part a single screenshot catches at random.

When a move stops being noise

The category does move. It moves slowly enough that a week is about the shortest window in which it shows:

By ten days, 21% and 49% of moves on ChatGPT and AI Mode beat the noise, and the median move was 1.55 and 2.65 points.¹

The most-named brand in a product line changed 4 of 54 times from one weekly wave to the next, and three of the four new leaders were ahead by less than the reading’s margin of error.¹

In a consumer finance software category we also measure, 0 of 14 ChatGPT moves beat the noise over ten days,⁶ and AI Mode’s new category leader was ahead of third place by 2.3 points, a tie inside the margin.⁶

Leon Claassen

Senior GTM Consultant @ Empact Partners

When AI visibility drops, check the instrument first, then the sample, then the engine, and only then the work. In the software category we measure, a brand’s share of answers moves about one point in four days, and most of those moves are chance. A team that reports each one as news learns to react to noise. A move earns its line in the report once it holds on the same questions long enough to rule the draw out.

Small Question Sets and Changed Instruments Manufacture Moves

The size of the question set decides how much of that noise lands in the number. We took every brand at 5% visibility or more in our category and re-measured it on random subsets of its product line’s questions:

On 25 questions, one reading in ten landed 11.8 points or more from the full-line figure.¹

On 100 questions, 5.7 points; on 200, 3.8.¹

The error falls with the square root of the sample, so the first hundred questions buy most of the precision and every hundred after that buys less.

Bars: the median gap. Line: the gap one reading in ten reaches or exceeds. One B2B software category, all five waves.

Otterly’s tracking gives the same shape on another ruler: 90% of brand-coverage readings fell between 39% and 71% on 10 prompts, and between 51% and 60% on 100.²¹

Claude shows it inside our own data. With about 100 answers per product line, a brand’s share of Claude’s answers moved a median 3 points in ten days, and 2.7% of those moves beat the noise.¹ The moves looked bigger because the sample was smaller.

The instrument moves too

The largest move between two waves anywhere in our category’s data is not a market event:

Between our 22 August and 2 September waves, the average Google AI Mode answer went from 13.92 to 3.36 citations, because the scraping vendor changed how it reads AI Mode.¹

The same week, brand visibility on AI Mode moved a median 5.7 points, up to 35.65, and 73% of moves beat the noise.¹

15.5% to 23.8% of mentioned brands overlapped when the same prompts were collected from the interface a user sees and from the API.²⁴

36% of apparent drop-offs on ChatGPT were the same company under a different name the next day.⁷

Nothing on a dashboard labels a week like that. Everything simply moves at once, on one engine, in both directions.

An Empact report measures visibility on hundreds of questions per category as a rolling baseline, reads each move against the category’s normal weekly range, and logs changes to the collection beside the numbers. That is how we measure AI search visibility for partners, and why our own AI Mode break reads as the vendor’s change rather than anybody’s result.

What Moves the Picture for Real: Releases, Retrieval and Google’s Dial

Take the draw, the sample and the instrument out, and what remains are the slow moves, the ones that change the pool itself. Almost all of them come from the engines themselves.

ChatGPT changes how often it looks things up

Whether ChatGPT searches the web before it answers decides which pages can reach the answer at all, and in 2026 that switch moved sharply, even inside a single model version:

37.3% to 48.7% of ChatGPT runs returned sources in Qvery’s June vertical studies, and 93.83% to 98.88% in August, across fintech, automotive and hotel questions.⁴

GPT-5.5’s answers carrying no citation went from 81% to 21% to 0.6% in May, June and August 2026, on the same model version throughout.²⁵

Qvery’s August question sets were narrower than its June ones, and narrower questions may trigger a search more often, so its new SaaS study reports the gap without explaining it: “Treat any trigger rate you measure as a reading taken on a date, not a fixed property of the engine.”

DataForSEO re-asks the same question set with every release, and its series shows the shift happening inside a single model version. A quarter-on-quarter comparison that crosses a change like this needs a note beside it.

And what it reads when it does

Qvery’s March 2026 citation trends caught ChatGPT changing how it builds answers inside a single month:

Per response, in Qvery’s tracking	February 2026	March 2026
ChatGPT citations	16.93	9.76
Google AI Mode citations	19.49	19.62
ChatGPT unique domains	21.2	11.7
Google AI Mode unique domains	13.2	13.3
ChatGPT brand mentions	7.22	6.32

The citations fell by about 42% while the brand mentions fell by an eighth.²⁶ seoClarity measured the same event from outside: the share of US ChatGPT responses with no citation at all went from 28% to 48% in March, before citations rebounded in May.²⁷ The mix of sources moved on its own schedule too:

ChatGPT’s citations to YouTube, LinkedIn, Facebook and Instagram fell 91% to 97% from January to April 2026 in Qvery’s social media study, while Google AI Mode’s YouTube citations grew 5.2 times.²⁸

88% fewer Reddit citations in ChatGPT ten days after GPT-5.6 became its default model in August 2026,²⁹ as the sources it retrieved per chat rose from 12.48 to 25.85.³⁰

How often ChatGPT’s citations lined up with Google’s results rose from 12% to 33% between April and July 2025, while the share lining up with Bing’s fell from 26% to 8%.³¹

52% less referral traffic from ChatGPT between 21 July and 20 August 2025, as its citations consolidated on a few answer sites, a shift that began weeks before GPT-5 shipped.³²

Google turns its own dial

Google also moves the surfaces themselves, on a schedule no page owner sets. Our panel asks AI Mode directly and cannot see how often AI Overviews appear, so these readings come from outside it:

42.4% of domains cited in AI Overviews before Google made Gemini 3 its default model no longer appeared after it.³³

Over 250 product launches shipped within AI Mode and AI Overviews in a single quarter, by Google’s own count.³⁴

6.49% to 24.61% to 15.69% of keywords triggered an AI Overview in January, July and November 2025.³⁵

23% to 47% to 34% of US Google searches showed one in September 2025, January 2026 and February 2026.³⁶

In Conductor’s data every one of the eleven industries swung between a low and a peak inside those six months, the narrowest from 13% to 29.3% of searches.³⁶

Spread chart of the lowest and highest monthly share of US Google searches showing an AI Overview in each of eleven industries between September 2025 and February 2026: Communication Services 26.9% to 63%, Health Care 31.6% to 56.6%, Financials 25.9% to 56.1%, Information Technology 25.3% to 52.4%, Energy 21% to 50.4%, Consumer Staples 25.2% to 45.2%, Industrials 19.7% to 44.6%, Utilities 19.2% to 39.5%, Materials 17.7% to 37.6%, Consumer Discretionary 17.3% to 36.5% and Real Estate 13% to 29.3%. — 274,524,214 US Google searches in 11 industries. Each bar runs from an industry’s lowest month to its highest in the window.

Leon Claassen

Senior GTM Consultant @ Empact Partners

Every one of those shifts reached the answers brands are measured on without anybody publishing a word differently. So a visibility report needs a change log beside the numbers: model releases, search-trigger changes, Google’s coverage moves, changes to the collection itself. When a brand’s visibility moves in the same week as an entry in that log, the entry gets investigated before anyone credits or blames the work.

More of the Buying Now Happens in the Engines That Move Most

None of this would matter much if the engines were a side channel. The searches are moving, less evenly than the headlines say, and fastest among the people a B2B software company sells to.

Where the searches are going

ChatGPT’s weekly audience grew about tenfold in under three years, by OpenAI’s own counts: more than 800 million a week by December 2025,³⁷ more than 900 million by February 2026,³⁸ and past a billion by August 2026.³⁹

Line chart of ChatGPT weekly active users as OpenAI stated them: more than 100 million a year after launch in November 2023, almost 350 million in November 2024, more than 700 million in July 2025 and more than 1 billion in August 2026. — Each point is the figure OpenAI published at that date, most of them floors (“more than”); the 2023 and 2024 points come from OpenAI’s usage paper. OpenAI also stated more than 800 million in December 2025 and more than 900 million in February 2026.

14% to 24% of ChatGPT use was seeking information, July 2024 against July 2025, which OpenAI’s own researchers call “a very close substitute for web search.”⁴⁰

1 billion monthly active users on Google AI Mode since its global expansion in October 2025.⁴¹

ChatGPT’s share of visits to generative AI websites fell from 76% to 53% between June 2025 and May 2026, while Gemini’s rose from under 9% to around 27% or 28%.⁴²

3.2% of US desktop searches across 41 major sites went to AI tools in the last quarter of 2025, against about 80% on traditional search engines.⁴³

So most searches have not left Google, and the ones that did are not all going to one engine. What moved fastest is the question that used to start a search, and the engines taking it are the ones whose answers move most, trading audience with each other as they go.

Why buyers move, and what it costs the click

B2B software buyers moved fastest, and the move reaches their shortlists:

B2B software buyers who start their research in a chatbot more often than in Google went from 29% to 51% in eleven months.⁵

53% of buyers find research with an AI chatbot more productive than traditional search, up from 36% seven months earlier.⁵

84% of B2B SaaS CMOs used AI tools for vendor discovery in January 2026, against 24% a year earlier, in a survey of 101 of them.⁴⁴

69% of buyers chose a different software vendor than they had planned because of a chatbot’s guidance.⁵

85% to 95% of the time, B2B buyers purchase from the shortlist they already hold before they talk to a seller.⁴⁵

Consumers give the reason in plainer terms: 53% of users choose a chatbot over a search engine to untangle a complex topic, 47% for step-by-step instructions and 45% to brainstorm.⁴⁶ The chatbot does the synthesis a results page leaves to the reader.

The click that used to carry a buyer to a vendor’s page is going with it. Google users clicked a traditional result on 8% of visits when an AI summary appeared, against 15% when none did.⁴⁷ The top-ranking page gets a 58% lower click-through rate when an AI Overview sits above it.⁴⁸

So the shortlist a buyer brings to a first sales call is increasingly drawn in one of these engines, from the pool this page has been measuring. A GEO partnership with us starts with an audit of that pool in the partner’s category: which brands the engines name, from which pages, and how steady each engine is, before any content or outreach is planned.

Read the Band Before Anyone Takes Credit

The line we read by: in a software category, a brand’s share of AI answers that moves by less than two points in a week, on a thousand questions, has not moved. A move that holds in the same direction for three weekly waves, on the same questions, with nothing in the change log that week, has. Everything in between is a reading to keep watching.

Reading moves this way is also why a GEO partnership with us is judged over quarters. Getting named across the pages the engines rotate through moves the pool slowly, and the engines move it faster, both ways, so a single month is weather and a quarter is climate. The standing report a partner gets from us shows the rolling baseline, never a screenshot.

If your AI visibility swings and nobody can tell you which week was the real one, book a call with us, and we will work out together whether this is work for us.

Sources

Empact Partners, Empact Panel, category A: one B2B software category, nine product lines; the same 9,000 unbranded buying questions on ChatGPT and Google AI Mode (900 on Claude) on 2, 6, 13 and 23 September 2026, and a subset of them on 22 August; every answer, citation and named brand, each question matched to itself across waves; read 28 September 2026.
SparkToro, “NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility”, 2026. sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility
SISTRIX, “AI Citation drift: How stable are sources in AI search results?”, 2026. sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results
Qvery, “First AI Recommendations for a Brand-New SaaS: What the Answers Had in Common”, 2026. qvery.ai/blog/new-saas-first-ai-recommendations
G2, “New G2 Research: Half of B2B Software Buyers Now Start Their Research With AI Chatbots”, release of the report “The Answer Economy”, 2026. prnewswire.com/news-releases/new-g2-research-half-of-b2b-software-buyers-now-start-their-research-with-ai-chatbots-302742807.html
Empact Partners, Empact Panel, category B: one consumer finance software category, US; the same 1,000 unbranded buying questions on ChatGPT and Google AI Mode on 15 and 25 September 2026; read 28 September 2026.
Detailed.com, “How Volatile Are AI Responses? (70K Answer, 28 Day Study)”, 2026. detailed.com/ai-volatility
AirOps, “The 2026 State of AI Search: How Modern Brands Stay Visible”, 2025. airops.com/report/the-2026-state-of-ai-search
Mike Sonders, “What repeated ChatGPT runs reveal about brand visibility”, Search Engine Land, 2026. searchengineland.com/repeated-chatgpt-runs-brand-visibility-468552
Qvery, “How To Measure Ecommerce Share Of Voice”, 2026. qvery.ai/blog/measure-ecommerce-ai-share-of-voice
OpenAI Help Center, “Searching the web with ChatGPT”, read 28 September 2026. help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
Google Search Central, “AI features and your website”, documentation, 2025. developers.google.com/search/docs/appearance/ai-features
Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference”, 2025. thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference
SE Ranking, “How Volatile Are AI Mode Results? 2025 Local Search Test”, 2025. seranking.com/blog/ai-mode-volatility-test
SE Ranking, “AI Mode Research: Sources, Volatility, & Differences between AIO and Organic Search”, 2025. seranking.com/blog/ai-mode-research
Radyant, “Do personas change what AI recommends? Here is what 17,929 chats revealed”, 2026. radyant.io/research/persona-study
Profound, “AI Search Volatility: Why AI search results keep changing”, 2025. tryprofound.com/blog/ai-search-volatility
Ahrefs, “AI Overviews Change Every 2 Days (But Never Change Their Mind)”, 2025. ahrefs.com/blog/ai-overview-change
Semrush, “Exploring URL Volatility in Google’s AI Overviews”, 2024. semrush.com/blog/url-volatility-ai-overviews
Authoritas, “AI Overviews & SERP Volatility: Research into how Google’s Search Results Change”, 2025. authoritas.com/blog/serp-organic-and-ai-overview-volatility-research
Otterly.AI, “AI Search Visibility: How Stable Are Brand Mentions and Citations?”, 2026. otterly.ai/blog/ai-search-visibility-stability
Ahrefs, “Are AI Mode and AI Overviews Just Different Versions of the Same Answer? (730K Responses Studied)”, 2025. ahrefs.com/blog/ai-overviews-vs-ai-mode
Qvery, “Have AI Engines Already Settled Your SaaS Category’s Top 5? Check Before You Try To Break In”, 2026. qvery.ai/blog/break-into-saas-category-ai-shortlist
Surfer, “Why You Cannot Track AI Visibility Through an API: 2026 Data From ChatGPT, Perplexity, Gemini, AI Mode, and AI Overviews”, 2026. surferseo.com/blog/llm-scraped-ai-answers-vs-api-results
Denis Pavlovsky (DataForSEO), LinkedIn post on ChatGPT citation rates across model versions, 9 September 2026. linkedin.com/posts/denis-pavlovsky-abb38926a_gpt-sitation-shift-2026-activity-7503418723240853505-qWX4
Qvery, “AI Search Citation Trends: March 2026”, 2026. qvery.ai/blog/ai-search-citation-trends-march-2026
seoClarity, “Tracking the Decline of ChatGPT’s Citations: A Global Trend Analysis”, 2026. seoclarity.net/chatgpt-citation-decline-analysis
Qvery, “Social Media in AI Search: ChatGPT vs Google AI Mode”, 2026. qvery.ai/blog/social-media-ai-citations-statistics
Tomek Rudzki (Peec AI), LinkedIn post on ChatGPT’s citations ten days into GPT-5.6, 18 August 2026. linkedin.com/posts/tomekrudzki_since-8th-august-the-default-model-for-chatgpt-activity-7495535968205201409-Cmsf
Tomek Rudzki (Peec AI), LinkedIn post on ChatGPT’s sources before and after GPT-5.6, 20 August 2026. linkedin.com/posts/tomekrudzki_chatgpt-is-falling-out-of-love-with-listicles-activity-7496162756547670016-PUKC
Profound, “AI Search Shift: ChatGPT’s growing alignment with Google’s index”, 2025. tryprofound.com/blog/ai-search-shift
Josh Blyskal (Profound), LinkedIn post on ChatGPT referral traffic, 20 August 2025. linkedin.com/posts/joshua-blyskal_chatgpt-referral-traffic-is-down-52-since-activity-7364003556087005185-BaJ0
SE Ranking, “Gemini 3 Replaces 42% of Previously Cited Domains in AI Overviews”, 2026. seranking.com/blog/gemini-3-impact-on-ai-overviews
Alphabet, “Alphabet earnings, Q4 2025: CEO’s remarks”, 2026. blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q4-2025
Semrush, “Semrush Report: AI Overviews’ Impact on Search in 2025”, 2025. semrush.com/blog/semrush-ai-overviews-study
Conductor, “Google AIO Volatility by Industry: What 2026 Data Reveals”, 2026. conductor.com/academy/ai-overviews-industry-volatility-analysis
OpenAI, “The state of enterprise AI”, 2025. openai.com/index/the-state-of-enterprise-ai-2025-report
OpenAI, “Scaling AI for everyone”, 2026. openai.com/index/scaling-ai-for-everyone
OpenAI, “A milestone in expanding access to AI”, 2026. openai.com/index/expanding-access-to-ai-with-chatgpt-ads
Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman, “How People Use ChatGPT”, NBER Working Paper 34255, 2025. nber.org/papers/w34255
Alphabet, “Alphabet earnings call Q2 2026: Sundar Pichai remarks”, 2026. blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026
Similarweb, “AI Search Stats in 2026”, 2026. aisearch.similarweb.com/blog/gen-ai-stats
SparkToro, “New Research: Search Happens Everywhere; an Analysis of 41 Websites with Significant Search Activity”, 2026. sparktoro.com/blog/new-research-search-happens-everywhere-an-analysis-of-41-websites-with-significant-search-activity
Wynter, “How B2B SaaS CMOs Buy Software in 2026”, 2026. wynter.com/post/how-b2b-saas-cmos-buy-software-in-2026
6sense, “The B2B Buyer Experience Report for 2025”, 2025. 6sense.com/science-of-b2b/buyer-experience-report-2025
Bain & Company, “Why Consumers Choose an AI Chatbot over a Search Engine”, 2026. bain.com/insights/why-consumers-choose-an-ai-chatbot-over-a-search-engine-snap-chart
Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 2025. pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results
Ahrefs, “Update: AI Overviews Reduce Clicks by 58%”, 2026. ahrefs.com/blog/ai-overviews-reduce-clicks-update

87 AI Search Volatility Statistics

Contents

Highlights

Ask Again and a Different Answer Comes Back