AI search study · October 2026

We looked for AI traffic hiding in Direct. We found bots.

Ramp says 30 to 40% of its Direct traffic comes from AI answer engines. Kevin Indig tested that idea on 12 sites and couldn't find it. We repeated his test on 10 sites we work with, added stricter statistics, and linked AI citations to individual pages.

From 10 sites Tenpoint Labs works with · GA4 and Google Search Console, January 2025 to 18 October 2026 · Published 22 October 2026

+164%growth in total Direct sessions, August to September, across 9 sites32,147 to 84,797 sessions

−1.6%change in engaged Direct visits to deep content pages over the same month448 to 441 visits

0Direct sessions carrying any AI tag or referral marker0 of 489,920 sessions

0 of 15links between AI referrals and later traffic that survive placebo tests and correctionHolm-corrected permutation tests

The claim we tested

“30 to 40% of our Direct traffic comes from answer engines.”

Ramp's growth team, as reported in Growth Memo. Their evidence: Direct visits to long blog URLs nobody types, behaving like search visitors, and customers saying AI sent them. Kevin Indig's replication on 12 sites found Direct growth was bots, and no reliable link between AI visibility and later traffic.

At a glance

Same question, different sites, the same answer

	Growth Memo (Kevin Indig)	This study
Sites	12 GA4 properties, mostly B2B and US	10 GA4 properties, mostly B2B and US
Sites with Direct growth, Aug to Sep	8 of 12, some over 100%	4 of 9, one at +281%
Total against engaged Direct	Total +140.9%, engaged −1.0% (Jan to Sep, filtered pages)	Total +164%, engaged deep pages −1.6% (Aug to Sep)
Worst bot site	409,000 Direct sessions, 198 engaged	153,614 sessions from one browser setup, 91% of the site's Direct
Direct sessions with an AI tag	0.003%	0 of 489,920
AI and later traffic	Weekly AI visibility (share of voice, ChatGPT mention rate) against Direct and branded search; 632 tests, best r = 0.72, failed placebo and split-half checks	Weekly AI referral clicks against deep-page Direct, homepage Direct, Unassigned and branded search; 15 tests with trends removed, best r = 0.45, none survive correction
New in this study	None	Behaviour match, a ceiling on hidden AI, and cited against uncited pages

Finding 1 · Bots

Direct traffic grew. Real visitors didn't.

Across the 9 sites with August and September data, Direct sessions rose from 32,147 to 84,797, a +164% jump. Engaged Direct rose +15%. Engaged Direct visits to deep content pages, where hidden AI traffic should land, went from 448 to 441.

Almost all of the jump came from one site. Site A's Direct went from 19,106 to 72,830 in a month, mostly iPhone Safari sessions with a fixed 998-pixel screen height, varying widths and almost no engagement. Site B had an earlier wave: 153,614 sessions from one Chrome on Windows setup at 3840 × 2160, 94% of them from China, landing on help articles between January and April.

Only 4 of 9 sites grew Direct at all. Kevin saw 8 of 12. September isn't a universal AI month; it's whatever the bots did that month.

Bar chart, pooled across 9 sites, August to September 2026: total Direct +163.8%, engaged Direct +15.1%, engaged Direct to deep pages −1.6%.

Bar chart of the share of Direct sessions flagged as bots per site, from 91% on Site B and 44% on Site A down to 2.1% on Site I.

Monthly Direct sessions on Site A and Site B, the two sites with the largest bot waves, split into sessions flagged as bots and other Direct, with engaged sessions as a line.

Per-site bars running from the change in engaged Direct to the change in total Direct, August to September. Long bars show growth that was mostly unengaged traffic.

Show the table

Site	Direct Aug	Direct Sep	Total change	Engaged change	Deep engaged Aug	Deep engaged Sep
Site A	19,106	72,830	+281.2%	+37.8%	13	49
Site B	316	260	−17.7%	−16.7%	6	4
Site C	7,441	6,998	−6.0%	−1.7%	230	209
Site D	2,038	1,679	−17.6%	−10.2%	110	91
Site E	122	129	+5.7%	+7.7%	8	8
Site F	899	655	−27.1%	−50.7%	61	42
Site G	1,230	942	−23.4%	+75.2%	10	30
Site H	436	515	+18.1%	+12.6%	0	0
Site I	559	789	+41.1%	+7.7%	10	8

Finding 2 · AI traces

Not one Direct visit carried an AI fingerprint

We searched all 489,920 Direct sessions for landing URLs carrying a ChatGPT, Perplexity, Gemini, Copilot or Claude source tag. ChatGPT now adds utm_source=chatgpt.com to many links, so a share of AI visits should arrive tagged even when the referrer is lost. We found none. When AI tools pass a tag, GA4 files the visit under the AI source, not Direct.

The AI traffic that is visible is small. It ranges from 0.09% to 3.0% of sessions per site, and ChatGPT sends 66% of it.

Line chart of monthly AI referral sessions on six sites, rising from 59 to 114 between October 2025 and September 2026.

Share of AI referral sessions by assistant: ChatGPT 66%, Gemini 16%, Claude 11%. Below, AI referrals as a share of all sessions by site, from 3.00% on Site E to 0.09% on Site J.

Finding 3 · Behaviour

Deep-page Direct doesn't behave like AI visitors

Ramp's argument rests on behaviour: Direct visits to deep pages acted like search visitors, so they were probably people sent by AI. We ran the same comparison. On 6 of 8 sites, Direct visits to deep pages engaged far less often than AI referrals or organic visitors. They look like bots, previews and link checkers, not people.

Then we gave the hidden-AI idea every benefit of the doubt. Assume every engaged Direct visit to a deep page came from AI. Visible AI visitors only land on deep pages part of the time (4% to 82% of visits, depending on the site), so we scaled up for the hidden ones that would land elsewhere. Even then, hidden AI comes to 6.3% of all Direct traffic across the 6 sites with enough AI visits to scale. Ramp's figure is 30 to 40%.

One small site breaks the pattern. Site E's ceiling reaches 42%, because only 13% of its 48 AI visits landed on deep pages. With so few visits the scaling is fragile, but on that site the data can't rule Ramp's range out.

Device mix didn't separate them either. ChatGPT's apps can drop the referrer, which would push hidden AI toward phones. Instead, deep-page Direct was mostly desktop, as AI referrals were.

Engagement rate per site for Direct visits to deep pages, AI referrals and organic visits to deep pages, June to September 2026. On 6 of 8 sites, deep-page Direct engages far less than AI referrals.

Show the table

Site	Direct deep: sessions	engaged	AI referral: sessions	engaged	Organic deep: sessions	engaged
Site A	363	50%	134	31%	91	59%
Site B	407	12%	42	67%	149	50%
Site C	2,896	26%	481	59%	2,448	48%
Site D	2,431	16%	145	45%	7,199	45%
Site E	48	60%	48	52%	313	71%
Site F	713	20%	26	38%	244	63%
Site G	561	11%	49	35%	160	34%
Site H	–	–	–	–	–	–
Site I	348	6%	25	24%	48	44%
Site J	377	20%	–	–	74	61%

Ceiling for hidden AI traffic per site, compared with Ramp's estimate of 30 to 40%. Site E reaches 41.7% on only 48 AI visits; the other sites sit at 9.6% or below.

Site D by month: AI referral sessions rose, while engaged Direct visits to deep pages per engaged organic visit did not follow.

Finding 4 · Correlation

Strong correlations, mostly shared trends

We lined up weekly AI referrals against five outcomes: engaged Direct visits to deep pages, engaged Direct visits to the homepage, engaged sessions GA4 couldn't assign to any channel, branded search impressions, and new branded queries. That made 15 tests across 5 sites, with 58 to 93 weeks each.

On raw weekly counts the picture looks compelling: correlations up to 0.83. That is the trap Kevin described. Every series grew through 2026, so any two of them correlate. Once we compared week-to-week changes, none of the 15 links survived correction for testing 15 at once. Site A's deep-page Direct, at 0.74 raw, fell to 0.30.

One result is worth watching. On Site B, engaged homepage Direct rose and fell with AI referrals in the same week (r = 0.45), in both halves of the timeline, and the link held after controlling for organic, email, referral and social traffic. It misses the corrected threshold (p = 0.07), rests on about five AI visits a week, and moves in the same week rather than later, which fits a shared cause such as a campaign as well as it fits AI. The homepage is where an AI-influenced visitor who types the brand would land, so it's the place to keep measuring.

These tests could reliably detect a correlation of about 0.3 or more. A weaker effect could still exist.

Dot chart of raw weekly correlations against the same tests with trends removed, with a placebo threshold for each. Raw correlations look strong; detrended ones mostly fall below the placebo line.

Show the table

Test	Weeks	Raw r	Trend removed r	Lag (weeks)	Placebo 95th	p	Corrected p	First half	Second half	Detectable r
Site B · Engaged homepage Direct	61	0.54	0.45	0	0.35	0.004	0.07	0.50	0.42	0.35
Site A · Engaged deep-page Direct	93	0.74	0.30	5	0.29	0.033	0.47	0.42	0.19	0.30
Site D · Engaged homepage Direct	61	-0.21	0.37	0	0.36	0.042	0.55	0.52	0.06	0.35
Site C · New branded queries	61	0.09	0.32	6	0.35	0.096	1.00	0.41	0.27	0.37
Site E · Engaged homepage Direct	93	0.15	-0.22	1	0.28	0.219	1.00	-0.28	-0.18	0.29
Site B · Engaged deep-page Direct	61	0.69	0.26	0	0.36	0.295	1.00	0.40	0.14	0.35
Site D · Engaged deep-page Direct	61	0.83	0.24	0	0.35	0.363	1.00	0.27	0.15	0.35
Site B · Branded search impressions	61	0.11	-0.23	5	0.35	0.424	1.00	-0.17	-0.27	0.37
Site B · Engaged Unassigned	61	0.29	-0.22	3	0.36	0.454	1.00	-0.19	-0.27	0.36
Site A · Engaged homepage Direct	93	0.77	0.18	4	0.30	0.486	1.00	-0.06	0.37	0.29
Site A · Engaged Unassigned	93	0.36	0.17	2	0.28	0.551	1.00	0.16	0.17	0.29
Site D · Branded search impressions	58	0.05	0.18	0	0.36	0.708	1.00	0.36	0.01	0.36
Site E · Engaged deep-page Direct	93	0.15	-0.12	2	0.28	0.847	1.00	-0.03	-0.24	0.29
Site A · Branded search impressions	68	0.25	-0.14	4	0.33	0.891	1.00	-0.13	-0.16	0.35
Site C · Branded search impressions	68	0.40	0.11	0	0.32	0.968	1.00	0.21	0.08	0.34

Correlation by delay in weeks between AI referrals and engaged homepage Direct on Site B. The same-week link clears the placebo line but not the correction for 15 tests (corrected p = 0.07).

Site A weekly AI referral sessions and engaged deep-page Direct, each on its own scale. Both rise in 2026, which is all the raw 0.74 correlation picks up.

Finding 5 · Citations

Cited pages get more AI clicks, not more Direct

For Site C we had AI answer runs from September listing every page each engine cited. Of the 33 blog pages with enough organic traffic to compare, 15 were cited.

Holding organic traffic constant, cited pages received +82% more AI referral sessions (95% interval +4% to +217%). Citations do send clicks, and GA4 sees them. Engaged Direct on the same pages was +17% higher, with an interval running from −22% to +76%, so no detectable difference. If citations pushed people into Direct, this is where it would show.

Effect of being cited on blog pages with the same organic traffic, with 95% intervals: AI referral sessions +82%, engaged Direct sessions +17%.

Scatter plot of blog pages by organic and engaged Direct sessions, log scales. Cited pages sit along the same line as uncited ones.

What to do instead

Stop reading Direct as an AI signal

Filter bots before any trend

Break Direct down by browser, OS, screen and country. One setup with near-zero engagement is a bot wave, and it can outweigh a site's real traffic for months.

Track engaged sessions, not totals

Engaged Direct held steady while totals swung by hundreds of percent. Report the engaged figure, or the total will tell the bot story.

Measure AI visibility directly

Run a fixed set of buyer prompts across ChatGPT, Perplexity, Gemini and Google's AI answers, and record whether the brand is named and which pages are cited.

Ask people how they found you

Add AI assistants to the "how did you hear about us" options. It is the one signal that reaches the visitor who never clicks.

Connect it to revenue

Put visibility, self-reported discovery and pipeline on one monthly report. When they disagree, check whether the tracked prompts match how buyers actually research.

Distrust any correlation that isn't detrended

Two series that both rise in 2026 will correlate at 0.7 or more whatever the cause. Test week-to-week changes against shifted placebos.

Method

How we ran it

Sites and data

Ten sites Tenpoint Labs works with, anonymised as A to J: mostly B2B software, plus a marketing agency, a business services firm and a consumer app. Most of their traffic is from the US. GA4 sessions from January 2025 to 18 October 2026, and Search Console from June 2025, through the APIs.

Two further sites were dropped for too little data, and one because its tracking stopped in March 2026.
Site C changed its GA4 setup in May 2026, so its landing-page figures start in June.
Site G's tracking starts in May 2026, Site I's in June and Site J's in September.

Deep pages and bots

The hidden-AI ceiling divides engaged deep-page Direct by the share of visible AI referrals that land on deep pages, on sites with 40 or more AI visits from June to September.
Deep pages are URLs under content sections (blog, post, guides, compare and similar) whose last segment has three or more words or 25 or more characters. Homepages, pricing, sign-in, help centres, tools and campaign pages don't count.
A visit is flagged as a bot when its browser, OS, screen and device setup has 20 or more sessions in a day at under 5% engagement and more than five times its usual volume, or 200 or more sessions in a month at under 2% engagement.
GA4 keeps browser version for only two months on most of these sites, so we could not use it across the full period.

Statistics

Weekly series, first differences of log counts to remove shared trends, lags of 0 to 6 weeks.
The placebo test shuffles the AI series 5,000 times and gives every shuffle the same 7-lag search, so the real timing has to beat the best of its own misalignments.
Split-half checks, Holm correction across all 15 tests, and the smallest correlation each test could reliably detect (80% power), which was between 0.29 and 0.37.
Branded search counts queries of five words or fewer. Longer branded queries are mostly AI fan-out searches that Google logs as impressions. They were 1.4% of branded impressions or less on every tested site.
A test needed 30 or more weeks and 100 or more AI referral sessions.

What this can't rule out

Effects weaker than a correlation of about 0.3. A small, steady AI share of Direct would not show up.
AI influence through visibility alone. Our weekly measure was AI referral clicks; Kevin's was share of voice, which is closer to the question.
Hidden AI traffic in sessions with no landing page, which we treated as bots, and bots that GA4 counts as engaged.
The page comparison covers one site and 33 blog pages, at a single point in time.
No customer attribution data was available, which is the evidence Ramp relied on.

Replicates

Kevin Indig, “Does AI drive more direct traffic?”, Growth Memo, 19 October 2026.

Cite this study

Tenpoint Labs (2026). "Is AI Hiding in Direct?" 22 October 2026. tenpointlabs.com/research/ai-direct-traffic

More research

All research →

Study · Oct 2026

YouTube in AI answers

31%of Google AI Overviews cite a YouTube video

Read the study →Study · 18 Oct 2026

Perplexity ranking factors

89%of sources retrieved first get cited, against 29% at tenth

Read the study →

See how often AI answers name your brand

Get the free AI visibility audit