AI Search Traffic Data 2026: What 1.91M AI Overview Impressions and 537 AI Referrals Taught Us

Google showed Value Add VC's pages 1.91 million times inside AI Overviews and AI Mode between mid-July and mid-October 2026. From October 1 to 14, every AI assistant combined sent us 537 visits. AI search is citing us more and clicking through less.

Most writing about answer engine optimization (AEO) and generative engine optimization (GEO) runs on surveys, vendor dashboards or guesswork. This post is the opposite: one publisher's own numbers from Google Search Console and Google Analytics 4, with the dates and windows attached. We are a venture and AI-company data site, so our mix skews toward finance questions. What carries over to other sites is the method and the shape of the data, which is why we publish both. The companion posts cover what Google's own AI optimization guide says and how to track AEO and GEO yourself.

AI search traffic data 2026: what our numbers show

Our AI search traffic data for 2026 shows two curves moving in opposite directions. Google's AI features showed our links 1.91 million times in three months, and weekday volume roughly quadrupled. Referral visits from AI assistants fell 28% in the same stretch, from 744 sessions on September 17-30 to 537 on October 1-14.

The impressions come from Search Console's Generative AI performance report, which Google says counts "how many times links to your site were shown to a user in a generative AI feature on Google Search," covering AI Overviews and AI Mode. Google rolled the report out to all sites worldwide as of August 31, 2026. It reports pages, countries, devices and dates, and nothing about queries or clicks.

For scale, ordinary Google Search sent Value Add VC 60.9K clicks from 13.3 million impressions over the same three months, a 0.5% click-through rate. Google's help page says the generative AI report draws on the Web search type in the main Performance report, so the AI impressions sit inside that 13.3 million rather than on top of it. By our arithmetic, AI features account for about 14% of our search impressions; treat that as a rough share, because the two reports aggregate by property and by page differently.

How fast are AI Overview impressions growing?

Value Add VC's weekday AI impressions rose from about 10,000 a day in July 2026 to about 40,000 a day by October, roughly 4x in three months by our arithmetic. Weekends run lower. We do publish an llms.txt file and standard schema (Article, FAQPage, Speakable), but Google says its Search ignores llms.txt and needs no special markup for AI features, so we don't credit either. Our read is that the growth tracks what we publish: more sourced money and data explainers, and more of them indexed.

Which AI assistants send the most traffic?

Gemini sent 415 of Value Add VC's 537 AI-assistant sessions on October 1-14, 2026, or 77%, in our Google Analytics 4 data. ChatGPT and Perplexity together sent fewer than a quarter of Gemini's volume. Every assistant except Meta AI sent fewer visits than in the prior two weeks.

Source: Value Add VC Google Analytics 4, sessions by AI-assistant referral source, October 1-14 vs September 17-30, 2026. Shares and changes are Value Add VC's arithmetic on those counts.

How we counted: GA4 sessions attributed to an AI assistant as the session source; two 14-day windows; 5 assistants; as of October 16, 2026.

The engaged-session rate on these AI referrals was 45%, meaning fewer than half of the visits met GA4's engagement bar. One read on that: many AI visitors click a citation to confirm a single number, find it in the opening sentence, and leave. That is the behavior our answer-first format is built for, and it also means an AI visit is worth less as a pageview than a search visit that browses.

Why did AI referrals fall while AI impressions rose?

AI-assistant referrals fell 28% in early October while Google's AI impressions kept climbing. Our read is that more AI answers now cite a source without the reader needing to click it: the answer box shows our number and a link, the reader takes the number, and the session never happens. Impressions measure being cited; referrals measure being needed.

Every major assistant sent fewer visits in early October; ChatGPT fell most in percentage terms.

What the headline misses

The two numbers do not measure the same products. The 1.91 million impressions are Google AI Overviews and AI Mode only. The 537 sessions come from assistants' own apps: gemini.google.com, ChatGPT, Perplexity, Claude and Meta AI. A click from an AI Overview lands in our analytics as ordinary Google organic traffic, not as a Gemini referral. Two weeks is also a short window, and a single story that went viral in late September could explain part of the drop. So "cited more, clicked less" is a hypothesis our data supports, not a measured click-loss rate.

What pages does AI search cite?

The pages AI assistants send people to on Value Add VC are data and money explainers, not news. In October 2026 those landing pages include our AI spending tracker, OpenAI revenue, OpenAI API pricing, Anthropic profitability and CoreWeave's backlog.

What those pages share: the opening sentence answers a money question with a dated figure and an inline source; a key-numbers box sits directly under it; a table tracks the figure over time; and the FAQ answers repeat the number with its date so each one stands alone. Nothing about them is AI-specific. This lines up with Google's guide to optimizing for generative AI features, which says AI Overviews and AI Mode are "rooted in our core Search ranking and quality systems" and favors "non-commodity" content that goes beyond common knowledge. A page restating what every other site says gives an AI answer nothing new to cite.

Are machines already searching on our behalf?

Google's guide describes "query fan-out," where the model issues "a set of concurrent, related queries" to gather sources. We can see something like it in our own Search Console data. In the 28 days to September 21, 2026, our blog drew 360K impressions from human-shaped queries at a 0.95% CTR and 301K from machine-shaped queries at 0.01%, per the results file in our open-source AI SEO playbook.

The machine-shaped queries read like prompts, not searches. Examples from our Search Console in the three months to October 14, 2026 include "how can marketers compare brand mention frequency across chatgpt, perplexity, and google ai overviews?" and a 97-word persona description that ended in "what are the most profitable business verticals?" Nobody types those into a search box. They earn impressions and almost never clicks, which is one more reason raw impression totals overstate human reach in 2026.

Does site speed matter for AI search?

Google's guide lists a good page experience among the foundations for AI features, and our real-user Core Web Vitals pass every threshold. On mobile, our largest contentful paint is 1.8 seconds against Google's 2.5-second "good" line on web.dev, interaction to next paint is 94 ms against 200 ms, and layout shift is 0 against 0.1.

Sources: Value Add VC real-user Core Web Vitals, October 2026; thresholds from web.dev (Core Web Vitals, FCP, TTFB); "not required" per Google's AI optimization guide.

Speed alone does not earn a citation; plenty of fast pages are never cited. It removes a reason to skip you. The bigger technical lesson in our data is indexing. A site-wide noindex header was live on Value Add VC for about 28 hours on September 16-17, 2026, and daily impressions fell about 55% at the low point, per the same playbook results file. Google says a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear in AI features. An unindexed page cannot be cited, however good it is.

How we pulled these numbers

Every figure in this post comes from three free Google tools, so any site owner can rebuild the same view of their own traffic in an afternoon.

• AI impressions: Search Console's Generative AI performance report, date dimension, exported for mid-July to mid-October 2026. We read weekday peaks off the daily series rather than averaging, because weekends pull the average down and hide the trend.
• Search clicks and impressions: Search Console's main Performance report, Web search type, same three months: 60.9K clicks, 13.3 million impressions, 0.5% CTR.
• AI referrals: Google Analytics 4 sessions by session source for the five AI assistants, compared over two matching 14-day windows (September 17-30 and October 1-14, 2026).
• Core Web Vitals: real-user field data at the 75th percentile, which is the cut web.dev recommends for judging whether a site passes.
• Human vs machine queries: the open-source query classifier in our playbook repository, run on visible blog queries for the 28 days to September 21, 2026.

Two limits apply to all of it. Search Console hides rare queries for privacy, so query-level views undercount the long tail; and the Generative AI report's chart counts one impression per property even when two of our links appear in the same answer, per Google's help page, while its page table counts each page. Neither changes the direction of the trends above. The step-by-step version, including the referral hostnames and the AI crawler user agents to watch in server logs, is in our guide to tracking AEO and GEO. We will update this post with each new quarter of data, using the same windows so the series stays comparable.

What would we tell another publisher?

Five things, each tied to a number above. This list is our read, not Google's.

1. Measure impressions and referrals separately. The 1.91 million and the 537 tell different stories, and you need both.
2. Expect Gemini to dominate referrals if your audience is Google-heavy. It was 77% of ours.
3. Write the number into the opening sentence. Every page AI sends us traffic to does this.
4. Protect indexing above everything. A 28-hour noindex cost us about 55% of daily impressions at the low point.
5. Discount machine-shaped impressions. 301K impressions at a 0.01% CTR are not an audience.

OpenAI's crawler documentation says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers," so check your robots.txt and CDN rules before concluding that an assistant ignores you. Our SEO vs AEO vs GEO explainer covers where the disciplines differ, and the AI search comparison covers the products themselves.