Buyer Behavior Science
The Speed-to-Answer Standard: Four Decades of Delay Data and the AI-Answer Frontier Nobody Has Measured Yet
Every documented increment of delay carries a measurable cost, and the newest instant-answer layer, generative AI, is now the fastest-growing part of that story with the least independent measurement behind it.
Part of Choice Science in the Insights library.
Abstract
Response-time expectations have been compressing for over three decades, from Nielsen's 1993 attention thresholds through Google's own controlled speed experiments to today's near-instant AI-answer layer, and at every measured step, added delay carries a documented cost in abandonment, contact odds, or conversion. The newest layer, generative AI chatbots and AI-written search summaries, is being adopted faster than almost anything that preceded it and is increasingly satisfying people's need for an answer with no click-through to any source at all, yet unlike page-load speed or lead response time, it still has no independent, audited census of its true scale. That gap between fast, well-measured history and fast, poorly-measured present is the headline of this study.
The acceptable wait for a digital response has been compressing for more than thirty years, and the record is unusually well documented for once. Nielsen's 1993 thresholds set the original bar: 0.1 second to feel instant, 1.0 second before a person notices the pause, 10 seconds before attention leaves the task. Google's own 2009 experiment proved the stakes were real by deliberately adding 400 milliseconds to search results and watching usage drop, with part of that drop never fully recovering. Ecommerce expectations moved in lockstep: Forrester and Akamai found the acceptable page-load window narrowing from four seconds to two between 2006 and 2009, and Google's 2017 mobile research showed more than half of visits abandoned past three seconds even though the average mobile page still took 22 seconds to load. Then a new layer arrived on top of all of it. Pew found ChatGPT usage among US adults roughly doubled between 2023 and 2025, echoing the fast early adoption of voice assistants Pew measured back in 2017, and the Reuters Institute's 45-market survey shows AI chatbot use for news climbing globally, from 7% to 10% of weekly users in a single year. The market is not just getting faster, it is shifting toward a different kind of answer altogether, one assembled and synthesized rather than retrieved from a ranked list, and estimates of how much of search now opens that way (roughly 15.7% to 48% of queries depending on the vendor doing the counting) diverge sharply enough to show the shift is real but not yet measured with any single trusted yardstick.
Every rung of measured delay in this record maps to a real behavioral cost, and the costs compound the closer they sit to a purchase decision. A 400-millisecond search delay cost Google measurable usage that did not fully return. A three-second mobile load cost more than half of visits. Baymard's long-running checkout research puts average cart abandonment in the high 60s to low 70s percent range, with slow or complicated checkout consistently among the most cited reasons buyers give up on a purchase they had already decided to make. And nowhere is the cost sharper than in the moment right after someone raises their hand: the MIT/Kellogg lead-response research found the odds of reaching a web-form lead fall roughly a hundredfold, and the odds of qualifying it roughly twenty-onefold, if the callback comes at 30 minutes instead of 5. Buyers now describe their own patience in similarly urgent terms across every channel: HubSpot, Zendesk, and Salesforce surveys each find large majorities expecting near-immediate or real-time responses, though none of those three self-reported studies publishes the methodology needed to call the exact figures settled rather than directional. Underneath all of it sits a psychological pattern first named by David Maister in 1984: unoccupied, uncertain, and unexplained waits feel far longer than their clock time, which is why the perception of delay, not just its length, is what buyers are actually reacting to.
The end state of a decades-long compression toward instant answers is an answer that requires no click at all, and that is precisely where measurement runs out. The Reuters Institute found only 4% of people usually click through from an AI chatbot's answer back to its original source, compared with 19% from a search engine, and trust in AI-chatbot-delivered information sits at just 20% against 32% for search. Seer Interactive's tracking shows organic click-through on AI-Overview-triggering queries falling 61% over roughly fifteen months, even as that figure has been climbing back since early 2025. This redraws where clicks happen. It also redraws where a business's real visibility now lives: increasingly inside a synthesized answer a buyer never leaves, on a surface with no Nielsen-grade response-time standard and no single audited count of how often it even appears. That is precisely the frontier a Visibility Corpus is meant to chart, a live read of where a given market's attention actually sits across search, social, and the AI-answer layer, so that a business can see and act on ground truth long before any official census of the AI-answer layer exists to do it for them.
The data, in one read
The Tenth of a Second That Became Doctrine
In 1993, researcher Jakob Nielsen laid down three numbers that interaction designers still build around today: 0.1 second feels instant, a response inside 1.0 second keeps a person's flow of thought intact even though they notice the pause, and 10 seconds is roughly the outer limit before someone's attention drifts off the task entirely unless you give them something to watch, like a progress indicator. Nielsen Norman Group still runs this piece as a live reference more than three decades later, which tells you something on its own: nobody has found a reason to move the goalposts.
What is easy to miss is that these were never really about technology. They describe how a person's mind tracks the gap between an action and a reaction, and that gap has stayed remarkably stable even as the systems around it have gotten faster and stranger. A 2021 lab study out of Telefonica Research and Google, testing ten latency conditions between 337 milliseconds and nearly 13 seconds on real mobile search users, found people tolerate roughly 7 to 10 seconds on a phone before reporting real frustration, tenseness, and fatigue, a threshold about four times higher than earlier desktop research had found. Even the floor for patience depends on the device in someone's hand, and it has never been fixed.
0.1 second feels instant. Ten seconds is roughly where a person's attention leaves the task altogether. Neither number has moved in three decades.
What Happens When You Actually Test the Delay
Nielsen's numbers describe perception. Google went further and tested the money. In 2009, Google Research ran a controlled experiment: add exactly 400 milliseconds, less than half a second, to the time it takes search results to appear, and measure what happens. Searches per user fell about 0.44% in the first three weeks and 0.76% over the following three, and even after the delay was removed, usage stayed down about 0.21%. That last detail is the one worth sitting with. People did not just tolerate the slower version less, some of them changed their habits and never fully came back.
You will also see, almost everywhere delay gets discussed, a claim that every 100 milliseconds of added page load cost Amazon roughly 1% of sales, traced back to a 2006 remark relayed secondhand through a former Amazon engineer's blog. It is worth repeating because it is a good story, and worth flagging because Amazon has never published it anywhere you can check. We treat it here for what it is: credible industry lore, not a verified figure.
The Load-Time Bar Kept Dropping, Years Before AI
The compression did not start with generative AI. Trace the ecommerce data back and you can watch the acceptable wait shrink in real time. In 2009, Forrester Consulting, working for Akamai, surveyed just over a thousand online shoppers and found 47% now expected a page to load in two seconds or less, down from a four-second expectation in a comparable study only three years earlier, and 40% said they would leave a site that took more than three seconds to render.
Google's own 2017 mobile research found the average mobile page actually took 22 seconds to fully load, a real gap between what sites deliver and what people will wait for: 53% of mobile visits are abandoned once load time passes three seconds, and pages that loaded within five seconds saw 25% higher ad viewability, sessions 70% longer, and bounce rates 35% lower than pages that took 19 seconds. The pattern across a decade of testing is consistent even where the exact percentages come from different studies and different years: the number of seconds a business gets before someone leaves keeps getting smaller, not larger.
Checkout Is Where the Patience Runs Out for Good
Baymard Institute has run an ongoing meta-analysis of checkout abandonment since 2012, folding in dozens of individual studies, and its running average lands in the high 60s to low 70s percent range for shopping carts started and never finished. Complicated or slow checkout is consistently one of the most cited reasons shoppers give, alongside cost surprises and forced account creation, and Baymard's research estimates the recoverable revenue from fixing checkout friction at hundreds of billions of dollars industry-wide for large sites.
For a small or mid-size business, the number that matters is not the industry aggregate, it is your own checkout's step count and load time against a buyer who has already decided to spend money and is now deciding whether the process is worth the wait. That is a different kind of abandonment than someone leaving your homepage. They wanted to buy. The friction, not the interest, is what beat you.
The Lead Went Cold Before You Finished Your Coffee
The most dramatic number in the whole speed-to-answer literature is not about a webpage at all. It is about a phone call. A study run by InsideSales.com with a researcher from MIT/Kellogg School of Management analyzed more than 15,000 leads and over 100,000 dials across six companies and found that the odds of actually reaching someone who filled out a web form fall roughly 100-fold, and the odds of that lead qualifying fall roughly 21-fold, when the callback happens at 30 minutes after submission instead of 5 minutes.
That study is nearly two decades old now, from a world of phone callbacks, not synthesized answers. But it is the clearest, most independently rigorous evidence in this entire body of research that response speed is not a nice-to-have on the margins of a sale, it is close to the whole game at the exact moment someone raises their hand. Every minute a business takes to follow up after that moment is a minute spent handing the lead to whoever answers first, and increasingly, whoever answers first is not always a person anymore.
The odds of reaching a lead fall roughly 100-fold when the callback comes at 30 minutes instead of 5. That was true in 2007, on a telephone, before any of this was automated.
What Buyers Say They Expect (Read the Fine Print)
Three separate research houses, HubSpot, Zendesk, and Salesforce, have each fielded their own surveys asking customers directly what they expect from a response. HubSpot's live-chat research finds 90% of customers rate an immediate response as important or very important, with 60% defining immediate as 10 minutes or less, and 82% specifically expecting an immediate reply in live chat, even though the reported industry-average first response time runs somewhere between 46 seconds and 2 minutes. Zendesk's 2026 CX Trends survey finds 88% of customers now expect faster responses than they did a year earlier, 74% expect 24/7 availability, and 86% say responsiveness and accurate resolution strongly shape their purchase decisions. Salesforce's State of the Connected Customer research reports a majority of customers now expect real-time communication as standard, part of a broader finding that 80% weigh the experience a company provides as heavily as the product itself.
We are holding all three of these at contested rather than established, and we want to say plainly why: none of the three publishes full methodology, response rates, or raw data alongside the headline figures, which is the bar we hold every number in this study to before calling it settled. That does not make the direction wrong. Three independent vendors, surveying different customer populations in different years, converge on the same story: expectations for how fast a business should respond keep rising. Treat the trend as real and the exact percentage as directional.
Why a Bad Wait Feels Worse Than Its Own Clock Time
Long before anyone measured page load in milliseconds, Harvard Business School researcher David Maister laid out a set of propositions in 1984 that still explain most of what the digital data above is really describing. An unoccupied wait feels longer than an occupied one. An uncertain wait, where you do not know how much longer it will take, feels longer than a known, finite one. An unexplained wait feels longer than an explained one. None of that has anything to do with actual seconds.
This is why a spinner, a progress bar, or a message telling you an agent will join in two minutes exists at all: not to make anything faster, but to make the same wait feel shorter by removing uncertainty and giving the mind something to do. It is also why a fast but confusing checkout can feel slower than a slightly slower but clear one, and why the businesses winning on speed-to-answer are not always the ones with the fastest server, they are the ones that manage what the wait feels like as carefully as how long it actually is.
The Newest Instant Answer Is Arriving Fast, in the US and Everywhere Else
Layer the newest instant-answer surface on top of everything above and the pattern accelerates. Pew Research Center found the share of US adults who have used ChatGPT rose from about 18% in 2023 to 34% by early 2025, in a survey of 5,123 US adults fielded that February and March, essentially doubling in under two years. That is not the first time a quick-answer interface caught on fast: Pew found 46% of Americans were already using a voice-controlled digital assistant back in spring 2017, mostly on their phones, an early sign that attention was migrating toward interfaces that skip the search-results page entirely.
This is not a US-only story. The Reuters Institute for the Study of Journalism at Oxford surveys 45 markets each year for its Digital News Report, and the 2026 edition found weekly use of AI chatbots for news rose from 7% to 10% between 2025 and 2026, a real global signal even though it remains a smaller share of the population than search or social platforms.
How much of the search results page itself now opens with a synthesized answer instead of ranked links is a genuinely open question, and we want to show you the disagreement rather than smooth it over. Semrush, tracking more than 10 million keywords, found Google's AI Overviews appeared on about 6.5% of queries in January 2025, spiked to nearly 24.6% by July 2025, then settled around 15.7% by November. BrightEdge, measuring separately, put AI Overviews at roughly 48% of tracked queries by February 2026. Those two numbers, for roughly the same window, are two to three times apart. There is no single official or audited measure of how much of search has become an AI-written summary, the way there is a Nielsen standard for page response time or a Pew standard for chatbot adoption. That gap is not a footnote. It is the frontier this whole study is pointing at.
Two research firms measuring the same thing, in roughly the same window, land two to three times apart. There is no Nielsen threshold yet for the AI-answer layer. That is the current state of the measurement.
The Click That Increasingly Does Not Happen
If speed-to-answer has a logical endpoint, it is an answer that arrives without a click at all. The Reuters Institute's 2026 data gets at this directly: only 4% of people say they usually click through from an AI chatbot's answer to the original source, compared with 19% from a search engine and 17% from social media, and only 20% say they trust news delivered through an AI chatbot, against 32% for a search engine. People are getting their answer faster than ever, and verifying it less than ever.
Seer Interactive's analysis of search queries that trigger a Google AI Overview, spanning June 2024 through September 2025, found organic click-through on those queries fell 61%, from 1.76% down to 0.61%, though the same data shows AI Overview click-through has been climbing again since January 2025, a genuinely mixed and still-moving picture rather than a straight line down. We hold this one at contested for exactly that reason: real movement, direction still settling.
Put the whole arc together and the shape is clear even where individual numbers are still shifting. Response-time expectations compressed from four seconds to two, then to sub-second, over roughly two decades of measured behavior. The newest layer on top of that, a synthesized answer instead of a rendered page, is being adopted faster than almost anything that came before it, and it is satisfying people's need for an answer with almost no click-through back to the business, publication, or source behind it. That is the part of the speed-to-answer story nobody, not Nielsen, not Google, not Pew, has built an audited census for yet. It is where a business's actual visibility now lives, and it is not yet reliably measured by anyone, which is exactly the gap a Visibility Corpus, a real map of where a market's attention sits and how it is moving, is built to close.
The evidence, in numbers
Key findings, dated and sourced
-
Nielsen's three response-time thresholds anchor interaction-design practice: 0.1 second feels instantaneous, 1.0 second is the limit before users notice a delay while their flow of thought stays intact, and 10 seconds is the limit for holding attention on a task before users need explicit feedback such as a progress indicator.
established Nielsen Norman Group, Response Time Limits: Article by Jakob Nielsen (1993 (originating in Nielsen's book Usability Engineering); maintained as a current reference)
-
Maister's foundational waiting-line propositions establish that perceived wait duration, not actual duration, drives satisfaction and complaint behavior: unoccupied waits feel longer than occupied ones, uncertain waits feel longer than known finite waits, and unexplained waits feel longer than explained waits.
established David H. Maister / Harvard Business School, The Psychology of Waiting Lines (1984)
-
In a controlled Google Search experiment, delaying results pages by 400 milliseconds reduced searches-per-user by roughly 0.44% in the first three weeks and 0.76% in the following three weeks; the effect persisted at about -0.21% even after the delay was removed.
established Speed Matters (Google Research Blog) (2009-06)
-
A widely repeated industry figure holds that every 100 milliseconds of added page latency cost Amazon roughly 1% of sales, based on 2006 internal A/B tests described secondhand by a former Amazon engineer and later cited by Google's Marissa Mayer. Amazon has never published this figure officially; it is credible industry lore, not a verifiable public dataset.
contested Greg Linden (former Amazon engineer) / Google, Geeking with Greg blog (Greg Linden), recounting Marissa Mayer's Web 2.0 presentation (2006-11)
-
Analyzing mobile landing pages and ads, Google found the average mobile page took 22 seconds to fully load, yet 53% of mobile site visits are abandoned if pages take longer than 3 seconds; pages loading within 5 seconds saw 25% higher ad viewability, 70% longer average sessions, and 35% lower bounce rates than pages taking 19 seconds.
established Google (Think with Google), The Need for Mobile Speed (2017-02)
-
In a Forrester Consulting study of 1,048 online shoppers commissioned by Akamai, 47% of consumers expected a web page to load in two seconds or less, up from a four-second expectation in a comparable 2006 study, and 40% said they would abandon a site taking more than three seconds to render.
established Forrester Consulting (commissioned by Akamai Technologies), eCommerce Web Site Performance Today (2009-09)
-
Baymard Institute's ongoing meta-analysis of dozens of abandonment studies puts average documented shopping-cart abandonment in the high-60s to low-70s percent range; complicated or slow checkout is among the most-cited reasons, with a large estimated recoverable-revenue opportunity from better checkout design.
established Baymard Institute, E-Commerce Checkout Usability research program (ongoing research program since 2012, most recently updated 2026)
-
A study analyzing over 15,000 leads and 100,000+ dials across six companies found the odds of successfully contacting a lead fall roughly 100-fold, and the odds of qualifying it fall roughly 21-fold, when a callback happens 30 minutes after web-form submission versus 5 minutes after.
established InsideSales.com with MIT/Kellogg School of Management, Lead Response Management Study (2007)
-
HubSpot survey research finds 90% of customers rate an immediate response as important or very important when contacting customer service, with 60% defining 'immediate' as 10 minutes or less; for live chat, 82% of customers expect an immediate response, though the reported industry-average first response time is around 46 seconds to 2 minutes.
contested How Consumers Use Live Chat for Customer Service (HubSpot Research) (survey-based blog research, cited as current through 2026)
-
Zendesk's 2026 CX Trends survey found 88% of customers report expecting faster response times than they did a year earlier, alongside 74% expecting 24/7 availability and 86% saying responsiveness and accurate resolution strongly influence purchase decisions.
contested Zendesk CX Trends Report 2026
-
Salesforce's State of the Connected Customer research reports that a majority of customers now expect companies to communicate with them in real time, part of a broader finding that 80% of customers consider the experience a company provides as important as its products.
contested Salesforce, State of the Connected Customer, 6th Edition (2023-08)
-
The share of US adults who have used ChatGPT rose from about 18% in 2023 to 34% by early 2025, per a Pew survey of 5,123 US adults fielded February-March 2025.
established Pew Research Center, 34% of U.S. adults have used ChatGPT, about double the share in 2023 (2025-06-25)
-
46% of US adults reported using a voice-controlled digital assistant in spring 2017 (42% on smartphones), an early marker of population attention migrating to quick-answer voice interfaces.
established Pew Research Center, Nearly half of Americans use digital voice assistants, mostly on their smartphones (2017-12-12)
-
Across 45 surveyed markets, weekly use of AI chatbots for news rose from 7% in 2025 to 10% in 2026. Only 4% of respondents say they always or often click through from an AI chatbot's answer to the original source, compared with 19% from search engines and 17% from social media, and only 20% say they trust news delivered via AI chatbots, versus 32% for search engines.
established Reuters Institute for the Study of Journalism, University of Oxford, Digital News Report 2026 (2026-06)
-
Tracking more than 10 million keywords, Semrush found Google AI Overviews prevalence rose from about 6.5% of queries in January 2025 to a peak near 24.6% in July 2025, then settled around 15.7% by November 2025; separately, BrightEdge measured AI Overviews appearing on about 48% of tracked queries by February 2026. The divergence shows the AI-answer layer still has no single official or audited census of its true reach.
emerging Semrush / BrightEdge, as reported by Search Engine Land, Google AI Overviews surged in 2025, then pulled back: Data (reporting Semrush and BrightEdge figures) (2025-11 / 2026-02)
-
A Seer Interactive analysis spanning June 2024 to September 2025 found organic click-through rate on search queries that trigger a Google AI Overview fell 61%, from 1.76% to 0.61%, though the report also notes AI Overview CTR has been rising since January 2025, indicating a mixed and still-shifting picture rather than a uniform decline.
contested Seer Interactive data, as reported in Google AI Overviews surged in 2025, then pulled back (analysis period June 2024-Sept 2025)
-
A controlled lab study of 30 participants testing ten latency conditions from 337ms to 12,975ms found mobile web-search users tolerate delays of roughly 7-10 seconds before reporting significantly more negative affect, a tolerance threshold about four times higher than comparable prior desktop-search research had established.
established Ioannis Arapakis & Souneil Park (Telefonica Research), Martin Pielot (Google), Impact of Response Latency on User Behaviour in Mobile Web Search, CHIIR '21 (2021-03)
Learning outcomes
What this study teaches
- Every incremental second of delay you add, whether in page load, checkout, live chat, or lead follow-up, has documented, sourced costs behind it. Audit your slowest touchpoint first, not your prettiest one.
- The steepest cliff in this whole body of research is lead response: waiting 30 minutes instead of 5 can mean roughly a hundredfold worse odds of reaching the person who just raised their hand, long before any AI enters the picture.
- Buyers now describe their own expectations in near-immediate terms across live chat, customer service, and ecommerce. The surveys behind those numbers are self-reported and vendor-commissioned, so treat the magnitude with caution, but the rising direction across three independent vendors is worth acting on.
- Generative AI's instant-answer layer is growing fast, roughly doubling US adult usage in under two years, and it is increasingly satisfying people's need for an answer with almost no click-through back to any source. That means your existing traffic and conversion numbers likely understate how much of your market's attention has already moved to a surface you cannot see in your own analytics.
- The next place to invest is knowing whether your business is even present in the answer being given, not just how fast your pages load. That is the specific gap a Visibility Corpus is built to read.
Honest limits
What this does not yet settle
- No independent, audited census exists yet for the AI-answer layer's true scale: prevalence and click-impact estimates for AI Overviews diverge by roughly 2-3x for similar time periods across vendors, and there is no equivalent of a Nielsen or Pew official measurement standard for generative-answer surfaces the way there is for page-load speed or chatbot adoption.
- No large-scale, recent, peer-reviewed replication of the classic latency-revenue findings (the Amazon anecdote, Google's 400ms experiment) exists for today's generative-AI chat interfaces, where the object of the wait is a synthesized answer rather than a rendered page. This is a genuine frontier with no rock-solid figure yet.
- No public study isolates the speed of the answer as the specific, measured driver of AI-chatbot or AI Overview adoption versus other drivers such as convenience, single-source synthesis, novelty, or trust. Usage and trust levels are documented; the causal weight of speed itself is not.
- The vendor-commissioned expectation surveys cited here (Salesforce, Zendesk, HubSpot) report what customers say they expect, not independently measured or revealed behavior. None published full methodology, response rates, or raw data, so the gap between stated expectation and actual observed behavior in these specific studies cannot be independently verified.
- No dated, sourced figure was found quantifying the incremental revenue or retention cost of a slow first response specifically within AI-agent or AI-chatbot customer service, as distinct from the well-studied human live-chat and lead-response literature.
This is a synthesis of dated, attributed evidence, not a census. The AI-answer layer in particular has no independent, Nielsen-grade measurement yet, so readings of it are directional and named as a frontier, never presented as settled.
Straight answers
Frequently asked questions
Is the pressure to respond faster actually real, or just a feeling?
It's real and it's decades deep. Nielsen set the original benchmarks back in 1993 (0.1 second feels instant, 10 seconds is roughly where attention leaves the task), and every layer added since has compressed the acceptable wait further: page-load expectations moved from four seconds to two between 2006 and 2009, and more than half of mobile visits abandon past three seconds. Most recently, Pew found ChatGPT usage among US adults roughly doubled between 2023 and 2025, and the Reuters Institute found weekly AI-chatbot use for news climbing globally from 7% to 10% in a single year. The trend line is one of the most consistently documented patterns in the study.
Does a slow response actually cost a business money, or is that just industry lore?
Some of it is proven, some of it isn't, and the study is careful to separate the two. Google's own controlled 2009 experiment is the real one: adding just 400 milliseconds of delay cut searches per user by roughly 0.44 to 0.76%, and usage never fully recovered even after the delay was removed. The famous claim that every 100 milliseconds cost Amazon 1% of sales, on the other hand, traces back to a secondhand 2006 blog post and has never been published by Amazon itself, so the study flags it as credible industry lore rather than a verified figure.
How fast do I actually need to follow up with a new lead?
Faster than most businesses assume. The MIT/Kellogg School of Management study, run with InsideSales.com across more than 15,000 leads and 100,000 dials, found the odds of reaching a web-form lead fall roughly a hundredfold, and the odds of qualifying it roughly twenty-onefold, if the callback comes at 30 minutes instead of 5. It's an older study, from the phone-callback era rather than today's chat and AI tools, but it remains the most rigorous evidence in the whole body of research that response speed can be close to the whole game at the moment someone raises their hand.
If someone gets an answer from ChatGPT or an AI Overview, do they still visit my website?
Usually not. The Reuters Institute's 2026 data found only 4% of people usually click through from an AI chatbot's answer back to the original source, compared with 19% from a search engine, and trust in AI-chatbot-delivered information sits at just 20% against 32% for search. Separately, Seer Interactive tracked organic click-through on Google AI-Overview-triggering queries falling 61% over about fifteen months, though that figure has been climbing back since early 2025, so the study treats it as a real but still-moving trend rather than a settled one.
How much of Google search now shows an AI-written summary instead of a normal results page?
Nobody agrees, and the study is upfront that this is the real gap in the evidence. Semrush, tracking more than 10 million keywords, measured AI Overviews at about 6.5% of queries in January 2025, spiking to nearly 24.6% by July 2025, then settling near 15.7% by November. BrightEdge, measuring separately over roughly the same window, put the figure at about 48%. That's a two-to-threefold difference between two research firms measuring the same thing, and unlike Nielsen's response-time standard or Pew's chatbot-adoption numbers, there's no single official or audited census of how much of search has shifted to a synthesized answer.
So what should a small business actually do with all this?
Audit your slowest real touchpoint first, not your prettiest one, since every documented step of added delay in this study, from page load to checkout to lead callback, carries a measured cost. Then recognize that the newest and fastest-growing part of that story, the AI-answer layer, is also the part with the least independent measurement behind it, which means your normal analytics likely understate how much of your market's attention has already moved somewhere you can't see. That's the specific gap a Visibility Corpus is built to read: a live map of where a market's attention actually sits across search, social, and the AI-answer layer, so a business can act on it before any official census exists.
Provenance
References
- Nielsen Norman Group, "Response Time Limits: Article by Jakob Nielsen" (1993) https://www.nngroup.com/articles/response-times-3-important-limits/
- David H. Maister / Harvard Business School, "The Psychology of Waiting Lines" (1984) https://www.hbs.edu/faculty/Pages/item.aspx?num=21299
- Google Research, "Speed Matters" (2009) https://research.google/blog/speed-matters/
- Greg Linden, "Geeking with Greg" blog, recounting Marissa Mayer's Web 2.0 remarks (2006) http://glinden.blogspot.com/2006/11/marissa-mayer-at-web-20.html
- Google (Think with Google), "The Need for Mobile Speed" (2017) https://www.thinkwithgoogle.com/_qs/documents/2340/bc22e_The_Need_for_Mobile_Speed_-_FINAL_1.pdf
- Forrester Consulting for Akamai Technologies, "eCommerce Web Site Performance Today" (2009) https://www.ir.akamai.com/news-releases/news-release-details/akamai-reveals-2-seconds-new-threshold-acceptability-ecommerce
- Baymard Institute, "E-Commerce Checkout Usability" research program (since 2012) https://baymard.com/research/checkout-usability
- InsideSales.com with MIT/Kellogg School of Management, "Lead Response Management Study" (2007) https://www.leadresponsemanagement.org/lrm_study/
- HubSpot, "How Consumers Use Live Chat for Customer Service" https://blog.hubspot.com/service/live-chat-consumer-behavior
- Zendesk, "CX Trends Report 2026" https://cxtrends.zendesk.com/
- Salesforce, "State of the Connected Customer," 6th Edition (2023) https://www.salesforce.com/content/dam/web/en_us/www/documents/research/State-of-the-Connected-Customer.pdf
- Pew Research Center, "34% of U.S. adults have used ChatGPT, about double the share in 2023" (2025) https://www.pewresearch.org/short-reads/2025/06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/
- Pew Research Center, "Nearly half of Americans use digital voice assistants, mostly on their smartphones" (2017) https://www.pewresearch.org/short-reads/2017/12/12/nearly-half-of-americans-use-digital-voice-assistants-mostly-on-their-smartphones/
- Reuters Institute for the Study of Journalism, University of Oxford, "Digital News Report 2026" https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary
- Semrush, BrightEdge, and Seer Interactive data, as reported by Search Engine Land, "Google AI Overviews surged in 2025, then pulled back: Data" https://searchengineland.com/google-ai-overviews-surge-pullback-data-466314
- Ioannis Arapakis and Souneil Park (Telefonica Research), Martin Pielot (Google), "Impact of Response Latency on User Behaviour in Mobile Web Search," CHIIR '21 (2021) https://ar5iv.labs.arxiv.org/html/2101.09086
Every measured figure is dated to its capture and tagged with an evidence tier. Every cited work is real and locatable. Where an engine could not be captured this round, it is named as uncaptured, not estimated. Small-sample readings are labelled as directional.