AI Share of Search Traffic
Growth in visits does not tell you whether a channel is winning. If every channel grows at once, a tripling means the whole pie got bigger, not that AI took a slice of it. So we start with share.
Set against organic search, AI referrals climbed from 0.98% of the two combined in January 2025 to about 1.8% by the middle of that year, and have stayed there. The fourteen full months since June 2025 average 1.8%, moving between 1.4% and 2.1% with no trend.
In absolute terms AI is still small. In July 2026, across every account in our network, AI referrals brought 498,439 visits and organic search 26.1 million. These growth figures cover the whole network, a wider base than the 1,200 sites scored for fit later in the study.
AI Referral Growth
Referral traffic is traffic that arrives from a link somewhere else. When someone asks an AI assistant a question and clicks a source in the answer, that visit arrives as an AI referral.
AI referrals grew 3.1x between January 2025 and July 2026. Organic search grew 1.64x over the same window. Almost all of AI's extra growth came in the first half of 2025: since June 2025 the two have grown at about the same pace, 1.6x AI and 1.7x organic, which is why the share above is flat.
AI Traffic by Assistant
AI referral traffic is not one channel. It is eleven assistants with different answer formats, different citation habits and very different volumes.
ChatGPT sends 81% of all AI referral visits. Gemini sends 8%, Claude 5%, Perplexity 3% and Copilot 2%. The six smallest send 0.3% between them.
ICP Fit by Channel
ICP fit measures whether a visiting company matches the ideal customer profile of the site it landed on. The profile is not ours. Each customer defines Company Segments in the product, and a company counts as fit when it matches at least one of them.
AI referrals match at 15.8%, against 20.4% for organic search and 19.1% for paid search. Only organic social matches less often. The order of the chart follows how the visitor arrived: the two channels where an algorithm picked the destination sit at the bottom.
Some of that 4.6-point difference is not about quality. AI visitors come from different countries than organic ones, and we identify fewer companies in some countries than others. Holding country constant, the difference is 2.9 points.
ICP Fit by Assistant
Eleven AI assistants sent referral traffic during the window and five of them carry almost all of it. Splitting AI referrals this way matters because the assistants do not all fit alike.
ChatGPT matches at 16.3%, the lowest of the five. The other four sit between 17.5% and 21.0%, and none of them clearly leads: Copilot's 21.0% rests on 3,673 companies and Claude's 20.8% on 10,008, too close to separate. Because ChatGPT is 81% of assistant visits, it sets the assistant average on its own. Excluding it, the rest match at 18.9%, close to paid search at 19.1%.
Fit by Site Size
Every figure in this study is a network average across 1,200 sites of very different sizes. Whether it describes your site depends on how much each channel's fit rate moves when site size changes.
Split the cohort at rank 600 and AI referrals move 1.3 points, from 16.0% on the larger sites to 14.7% on the smaller ones. That is one of the smallest shifts of any channel. Direct moves 8.2 points and referral 4.1. Organic social is the only channel that fits better on smaller sites, by 3.4 points.
Fit Gap by Site
A network average does not describe any particular site. Yours could sit well above it or well below it, so the number that matters to you is the spread, not the average.
We measured 712 sites individually. AI fits worse than organic search on 486 of them and at least as well on 226. The median site sits 2.4 points below organic search, and the middle half runs from 7.7 below to 0.9 above.
Return Visits by Channel
Companies that arrive from AI visit 2.68 times on average, the fewest of the seven channels you can actually acquire through. Organic search brings 3.78 and email 5.52. A channel that sends companies once and never again is worth less than the same channel sending them every week.
What It Means for Next Year's Budget
“I think everyone is publishing the growth curve and missing the number that matters. AI's share of search traffic has been flat at 1.8% for a year. The companies it sends are close to paid search in quality, there just are not many of them yet.”
Methodology, Scope and What We Do Not Claim
Scope
This study covers the top 1,200 of our accounts with customer-defined Company Segments, ranked by July identified-lead volume. Between them they hold 2,960 segment definitions, 33,806,546 identified company visits and 3,681,009 company-source pairs.
Coverage: those 1,200 accounts carry 23.2% of clean identified-lead volume and 84.0% of the volume from accounts that define segments at all.
The AI bucket covers ChatGPT, Gemini, Claude, Perplexity and Copilot plus six smaller assistants: DeepSeek, Meta AI, Mistral, Poe, You.com and Grok, which send 0.3% of AI visits between them and are omitted from the assistant chart as directional only. Excluded everywhere: accounts flagged test or spam, referrer domains that impersonate AI assistants, and about 30 internal and demo accounts that duplicate leadfeeder.com traffic. Ads shown inside AI assistants are too few to measure and are left out of every comparison. The cohort still skews to customers who have defined an ICP in the product, so treat the rates as strong for that population and directional for all B2B websites.
Data sources
Visits, sources and visiting-company identity come from our visit stream. Ideal customer profiles come from the customers themselves: the Company Segments they defined in the product, not any definition of ours. Firmographics for the visiting companies come from our own enrichment.
Industry matching is code-based rather than label-based. Segments and firmographics both carry LinkedIn v2 industry codes, matched hierarchically, so a filter on 25 matches a company coded 25_54_15 and a localised label cannot corrupt the match.
Metrics and reference windows
There is one metric in this study: whether a visiting company matches the ICP of the site it landed on. It is reported two ways.
Companies that fit (formally, fit % of company-source pairs) is the figure to quote. A company-source pair is one row per customer per visiting company per source, so each company counts once per site per channel. This is the metric the study calls "companies that fit". Visits that fit is shown alongside it and rewards channels with repeat visits from the same company. Direct is the clearest example: 45.8% of its visits are ICP-fit but only 25.0% of the companies behind them.
One ratio in the return-visits section is derived from these counts rather than measured separately: visits per company is visits divided by companies. Direct's 19.72 visits per company is inflated by visitors who already know the site, so it is drawn off the axis rather than compared with the acquisition channels.
Windows differ by section, deliberately
Growth: January 2025 to 12 August 2026, on all clean accounts, a wider base than the fit cohort. Channel, site-size and robustness cuts: 14 July to 12 August 2026. Assistant cut: 1 April to 12 August 2026, because assistant-level volume needs the longer window. That is why all assistants together fit at 16.9% over the assistant window and AI referrals fit at 15.8% over the channel window. Rates are comparable across sections. Visit totals come from the growth data; the fit sections use rates, company counts and visits per company.
Seasonality
Every quality figure rests on a single four-week window, 14 July to 12 August 2026, in the European summer holiday period. We have not yet measured fit in other months. The quarterly re-run will show how much it moves.
Direct
Direct is 69% of all visits in the fit dataset and 37% of all matching companies. It is where a visit lands when its source was not recorded, so it mixes untagged paid and organic visits with people who typed the address in. Shares of matching companies are therefore reported on acquisition channels only, with Direct excluded: 1.4% for AI, or 0.9% with Direct counted in.
Scoring methodology
A visiting company is ICP-fit for a given website if it matches any of that customer's Company Segments on all dimensions that segment specifies. A dimension the segment leaves open is a wildcard. This mirrors how our visitor identification evaluates segments in the product, so the test is the one customers actually use.
There is no weighting and no normalisation. Every figure in this study is a plain rate: fit rows divided by total rows, for the cut in question.
Robustness checks
Standardisation
To separate channel quality from visitor geography, AI's per-country fit rates are reweighted to organic's country mix. Country cells under 30 companies are dropped and the rest renormalised, keeping 98% of AI volume. The crude 4.6-point gap becomes 2.9 points, with a 95% interval of plus or minus 0.4.
Replication
The primary robustness test is at site level: of the 712 sites with at least 10 AI and 10 organic companies of their own, 486 show AI below organic search and 226 do not. Under a sign test against an even split that is z = 9.7. The gap also holds in 28 of 31 countries with at least 200 AI-referred companies, in every region except LATAM, and in 19 of 21 disjoint account subsets ranked by volume. At a stricter 30-company floor the median is -2.2pp and 71% of sites sit below zero. Differences between assistants also narrow as site size falls: across the smaller half of the cohort the five majors sit within about two points of each other.
Measured twice
The larger 600 sites and the smaller 600 were measured separately. Holding country constant, they land at 2.9 and 2.5 points.
Missing data and what we do not claim
Every fit rate here is a lower bound, and none of the indexed growth lines should be annualised. Industry is known for 71 to 74% of visits in the major channels, and a company with an unknown industry can never satisfy an industry filter, so it scores not-fit. This depresses all channels roughly evenly: comparisons between channels hold, absolute levels are understated.
What we do not claim. We do not claim AI fits better than organic social: below rank 600 the two cannot be told apart (z = 1.2). We do not rank one assistant above another among the four majors other than ChatGPT: Copilot and Claude in particular are too close to separate. We no longer report that Claude fits twice as well as ChatGPT. That came from the 60-account first version, and on the full cohort the ratio is 1.3x. We do not report any figure for August 2026 as a whole, because the month is partial, 1 to 12 August only.
Cuts not yet available: visitor industry by channel, the country cut for the channels beyond AI, organic and paid, and assistant by country.
Updates
We re-run this study every quarter on the same method and the same fit test, and publish the new figures on this page. The next update, with data to the end of December 2026, is due in January 2027.
Privacy & GDPR
Aggregates only, at country, region and channel level. No visiting company is named anywhere in this study and none can be identified from it. No customer account is named or identifiable.
