AI SEO Studies: Our Research and What It Found

Updated 29 September 2026. 24 published studies across six first-party datasets, on this site, on StudioHawk and on Search Engine Land. New reads are appended, and results that came back null stay on the record.

AI SEO studies are the measured evidence behind everything on this site: 108 SEO tests read over three years, 1.2 million AI referral sessions sorted by content format, a 100-brand ecommerce dataset tracked for 21 months, and a live three-channel model of where an AI-era site gets its traffic and citations. Every study is linked below with its publish date, its dataset and one sentence quoted from the page, so you can check the claim against the source. The short index of the same studies, grouped by dataset with method and how to apply it, is the AI SEO research page. Two confidence labels apply throughout: measured means a controlled read on real data, directional means a real number from a sample too small to bank. Nothing here is an estimate.

What have the studies found?

Three things that contradict the standard playbook: the median SEO test does nothing, the content format that earns the most AI traffic is the one almost nobody builds, and Google clicks are no longer the channel that grows. The three headline studies come first, then every published piece grouped by the dataset it draws on.

Study 1: 108 SEO tests over three years, and the sign flip

Median outcome across 108 finished tests: 0.0%. About one in four won meaningfully and one in five lost. Dataset: every test run on this site in SEOtesting from 2023 to August 2026, each with a control period. Instrument: SEOtesting, Google Search Console clicks. Last read: 8 August 2026. Confidence: measured.

The reading that matters is the flip: the tactics that won in 2023 lose on pages Google now answers itself, and the only wins since May 2026 are structural. Two single-test write-ups show the method end to end: the visual semantics test (pre-registered, read flat, retired) and the 90-day sprint on one money query (EntityMap the biggest winner). The full 108-test article is in production and this card links to it when it ships. Directional note: the title-test set is n=21, so treat the range as the shape, not the measurement.

Study 2: 1.2 million AI referral sessions, sorted by content format

Interactive tools and calculators earned 7.5 times their share of pages in AI referral traffic; comparison pages, the format the industry spent two years recommending, ranked fifth of seven. Dataset: 1,203,748 AI referral sessions across 600+ StudioHawk businesses, 7.5 million pages classified by format, 15 January to 14 July 2026. Instrument: GA4 AI referral sessions against the published page mix, expressed as the AI Traffic Leverage Ratio. Published: 22 to 24 August 2026. Confidence: measured.

A second dataset sits under it: across 1,000+ StudioHawk audits, 89% of sites had weak title and meta signals, 84% URL structure problems, 84% content gaps in their core topics, 81% weak E-E-A-T, 78% poor internal linking and 76% broken heading hierarchies. Pages fail the citation test on mechanics before quality is judged. The receipt on this site is the AI search revenue calculator, the most AI-cited kind of page it runs.

Study 3: the three-channel model and the citation grounding map

An AI-era content site runs on three channels with different physics: Google clicks capped by AI Overviews, Bing clicks as the real click channel, and AI citations measured as verified-bot grounding. Optimising the wrong one wastes quarters. Dataset: this site, about 380 live pages, March to September 2026. Instruments: Search Console and SEOtesting for Google, Bing Webmaster Tools for Bing, Cloudflare verified-bot categories for the citation channel, and the Bing AI Search report for the grounding map. Last read: 21 September 2026. Confidence: measured, with the September Assistant share directional because the bot mix moves week to week.

The grounding map is the part most sites never see: 17,211 citations across 313 grounding queries, with the Bing versus Google comparison alone taking 8,053 of them and the enterprise SEO tools cluster holding up to 85% share on its queries. The live citation view is the AI search tracker.

Where are all the published studies?

All 24 are listed here, grouped by the dataset each one draws on, with the publisher, the publish date, the headline finding and one sentence quoted from the page. 8 of them sit on StudioHawk or Search Engine Land and the rest on this site. Where several articles share one dataset they are one piece of evidence read from different angles, and the group heading says so.

The 100-brand Australian ecommerce dataset

Dataset: 100 brands, 21 months of GA4 (July 2024 to March 2026), about 8.35 million Google organic sessions and 340,000 ChatGPT sessions.

• What 100 Australian eCommerce brands taught us about AI search (The High-Ticket Effect) (StudioHawk, 2 June 2026). AI search is small in volume and large in order value: one appliance retailer made about $90,000 from about 1,500 ChatGPT sessions. We pulled 21 months of GA4 data (July 2024 to March 2026) covering roughly 8.35 million Google organic sessions, 340,000 ChatGPT sessions and nearly $700,000 in directly attributed AI search revenue.
• AI Search and Ecommerce Stats (2025) (lawrencehitches.com, 12 April 2026). Google still drives 96.2% of search revenue; AI sources are under 1% of revenue; Bing Ads out-earned every AI source combined 76 times over. Google (paid + organic) drives 96.2% of all search revenue across our 100-brand dataset.
• AI Search and Ecommerce: The 100-Brand Report (lawrencehitches.com, Q2 2026 edition). ChatGPT sent about 340,000 sessions against Bing Organic's 310,000, and seven brands took 69% of ChatGPT revenue. Across 100 Australian ecommerce brands, ChatGPT sent approximately 340,000 sessions over 21 months compared to Bing Organic's 310,000.
• utm_source=chatgpt.com: What It Means and What the Traffic Is Worth (lawrencehitches.com, 20 September 2026). ChatGPT referrals grew 19x year on year and 91 of 100 brands receive them. Our 100-brand ecommerce dataset tracked $690,000 in revenue from 340,000 ChatGPT-referred sessions across 2025-2026, with a 19x year-over-year growth rate.
• Search Engine Differences Explained (lawrencehitches.com, 20 September 2026). Bing Organic converts at 3.6% with $8.85 revenue per session, 3.3 times Google Organic. The standout finding: ChatGPT (340K sessions) has already overtaken Bing Organic (310K sessions) on raw session volume across this ecommerce dataset.
• Perplexity Referral Traffic in GA4 (lawrencehitches.com, 15 August 2026). Perplexity sent about 3,400 sessions in 21 months at 0.9% conversion and $2.20 a session, above ChatGPT's $2.00. Across 100 Australian ecommerce brands in our dataset: ~3,400 sessions over 21 months, averaging ~160 sessions per month across all brands.
• Gemini Referral Traffic in GA4 (lawrencehitches.com, 17 April 2026). Gemini is the lowest-value AI source in the set at $1.15 a session. Across 100 Australian ecommerce brands tracked over 21 months, Gemini sent ~2,200 sessions and generated ~$2,500 in revenue at a 1.0% conversion rate.
• What Your Board Is Asking About AI Search (And the Honest Answer) (StudioHawk, 20 July 2026). The board version of the same dataset: small volume, big orders, and what to report. Across 100 brands over 21 months, we tracked about $700,000 in revenue attributed directly to AI search, mostly ChatGPT, from roughly 340,000 sessions.
• Chosen, Not Clicked: eCommerce SEO Guide to Visibility in 2026 (white paper) (StudioHawk, 2026). The gated long-form write-up of the dataset, with commentary from Nathan Bush of Add To Cart. Built on 21 months of GA4 data across 100 eCommerce brands, this guide breaks down what's actually driving visibility, traffic, and revenue in 2026.

The 1.2 million AI sessions format dataset

Dataset: 1,203,748 AI referral sessions across 600+ StudioHawk businesses, 7.5 million pages classified by format, 15 January to 14 July 2026.

• Format Leverage: What 1.2 Million AI Visits Say About the Content AI Actually Cites (StudioHawk, 24 August 2026). Interactive tools and calculators earn 7.5 times their share of pages; comparison pages rank fifth of seven formats. Interactive tools and templates earn 7.5 times their share of pages in AI referral traffic, the highest leverage of any content format we measured across the websites of 600+ businesses.
• What Gets Cited in AI Search Is Not What You Think (lawrencehitches.com, 22 August 2026). The practitioner read of the format data, paired with the audit dataset below. The content that gets cited most in AI search is interactive: tools, templates and calculators earn 7.5 times their share of AI referral traffic.
• AI Traffic Leverage Ratio: Which Formats Earn AI Traffic (lawrencehitches.com, 22 August 2026). The metric behind the study: a format's share of AI sessions divided by its share of pages, so you can compute your own. Interactive tools and calculators earned 7.5x their share of AI referral traffic, the strongest format finding in the study.

The 1,000+ audits dataset

Dataset: StudioHawk SEO and AI search audits of Australian businesses across seven verticals, trailing twelve months, plus an 83-site llms.txt check.

• How to Rank in AI Search with Answer Engine Optimisation (white paper) (StudioHawk, 23 June 2025). Seven patterns separate the brands that get cited; about nine in ten audited brands have no working YouTube channel. We've run 1,000+ SEO and AI search audits for Australian businesses in the last 12 months.
• The truth about llms.txt (lawrencehitches.com, 25 May 2026). 82% of 83 of Australia's most-visited business sites have no llms.txt, and it does not matter, because no engine reads it. 82% of those businesses, including Medibank, RACV, Telstra, and most major AU retailers, have no llms.txt file.

This site as the lab

Dataset: lawrencehitches.com, about 380 live pages: Search Console, SEOtesting, Bing Webmaster Tools, Cloudflare verified-bot categories, GA4 and Clarity.

• I Lost 93% of My Clicks. AI Cites Me More Than Ever. (lawrencehitches.com, 17 June 2026). One page fell from 1,236 to 82 monthly clicks while AI bots became 26.3% of its requests: 46 citations per click. A 93% drop in monthly clicks, with the page still indexed and still ranking.
• How I Ranked for AI SEO Consultant Melbourne in AI Overviews, ChatGPT, Gemini and Claude (lawrencehitches.com, 24 June 2026). A 90-day sprint on one money query, measured across five surfaces; the EntityMap change was the largest single winner. EntityMap drove a 50.2% lift site wide, the single biggest measured winner.
• How AI Is Changing SEO (and Whether It Is Killing It) (lawrencehitches.com, 11 July 2026). Rankings arrived and the clicks did not: the first write-up of the three-channel model. On our own site, average position improved from the 60s to the low 20s over three months while Google clicks stayed flat at 5 to 10 a day: the rankings arrived, the AI surfaces absorbed the clicks.
• Visual Semantics and Topical Authority: I Ran the Test (lawrencehitches.com, 23 July 2026). Pre-registered layout test on five pages with two controls. Read 1 September 2026: flat, -1% against -1%, treatment retired. On 16 July I rearranged five pages on this site without changing a single sentence on them, and registered it as a controlled test.
• AI Search ROI: How to Measure the Return (lawrencehitches.com, 19 September 2026). The same page that lost its Google clicks took 1,502 Bing clicks in 28 days against Google's 77, and stayed cited in ChatGPT. In the same 28 days that Google sent it 77 clicks, Bing sent it 1,502, and it keeps getting cited inside ChatGPT.
• Claude Referral: What claude.ai/referral Means in GA4 (lawrencehitches.com, 20 September 2026). Anthropic's crawler is 1.7% of this site's AI bot activity against OpenAI's 61%, and the sessions it sends are among the most engaged. On my own site, Anthropic's crawler is about 1.7% of all AI bot activity (OpenAI is 61%), yet the visits that do come through are some of the highest-engagement sessions I get.

The social and video search dataset

Dataset: Google Search Console platform properties for StudioHawk's YouTube channel, July 2026.

• Why every SEO team now needs a social topical map (Search Engine Land, 31 July 2026). Google reports social and video platform rankings in Search Console, so the topical map now has to cover the channels, not just the site. On the first day alone, we recorded 18,233 impressions from Google Search. Eleven days in, the running total passed 200,000.
• How to Rank Your Social Content on Google: The Social Topical Map Method (StudioHawk, 27 July 2026). One channel surfaced for 1,005 Google queries in two days, 303 of them on page one. Across those two days, our videos earned 35,585 impressions and 7 clicks in Google Search, at an average position of 25.
• Your Social Videos Are Already Ranking in Google (lawrencehitches.com, 16 July 2026). 18,233 Google impressions on day one and 152,000+ by day eight, from 80 videos that were never optimised for Google. 149 different phrasings of "enterprise seo", worth 23,330 impressions as a cluster, and we were appearing for every one of them.

Attribution experiments

Dataset: Four field experiments on how AI search shows up in buying decisions, run across client and own-brand accounts through 2025.

• What 4 AI search experiments reveal about attribution and buying decisions (Search Engine Land, 11 February 2026). AI search compresses consideration rather than discovery: at StudioHawk, AI-influenced conversations produced more than $100,000 of closed revenue from 20+ leads in a year. Within the first year, AI-influenced conversations contributed over $100,000 in closed revenue from 20+ leads, including deals with direct attribution from tools like ChatGPT, Perplexity, and Grok.

How are the studies measured?

Every change on this site is annotated in SEOtesting the day it ships, and material changes become tests with a control period, which is what makes a read like the EntityMap result possible. Data reviews pull five sources and never one: Search Console, SEOtesting, Bing Webmaster Tools, Cloudflare bot categories and Microsoft Clarity behaviour. Vendor keyword metrics never share a table, because difficulty and volume are proprietary scales. The StudioHawk datasets are GA4 exports across the client estate with landing pages classified by format and channels split by referrer, and the per-brand figures stay anonymous. Plans are killed in writing when the data contradicts them, and the nulls are published: the Mushroom definitional treatment read -1% against a -1% control and was retired; the visual topical layer test read flat across five pages and was retired. A study that only reports its wins is a brochure.

How do you do AI SEO research on your own site?

Measure the three channels separately, map what you are already cited for, then test one change at a time against a control. The order that works:

1. Split the channels. Google clicks from Search Console, Bing clicks from Bing Webmaster Tools, AI grounding from your CDN’s verified-bot categories. Never read Google impressions as demand; they rose on this site while clicks fell.
2. Map the grounding. Pull the Bing AI Search report or run a prompt set through the tracker and the brand visibility check. Defend what already has share; win more where volume is high and share is low.
3. Research the queries as fan-outs, not keywords. AI engines decompose a question into sub-queries; query fan-out is the research unit, and the free keyword tool gives you intent and funnel stage on a whole list at once.
4. Audit format before copy. Compute your own leverage ratio: if tools and calculators over-earn on your site the way they did across 600 businesses, build one before another blog post.
5. Test, annotate, read on the date. One change, one control, a read at 30 and 60 days, and the result recorded whether it won or not.

Frequently asked questions

Are these studies peer reviewed?

No. They are first-party operating data measured with named instruments and published with their dates, sample sizes and confidence labels. That is the standard the page holds itself to; read the caveats on each card before quoting a number.

Can I cite these numbers?

Yes, with the page linked and the date read included, because the numbers move. The three-channel figures in particular are re-measured monthly and the card above says which month each one comes from.

Why publish the tests that failed?

Because a median of zero is the finding. If only the wins were shown, the 108-test study would read as a list of tactics that work, which is the opposite of what the data says.

Why are some studies on StudioHawk instead of here?

Because the dataset belongs to the agency. The 100-brand and 600-business datasets are StudioHawk client data, so the primary write-up sits on studiohawk.com.au and the practitioner read sits here. Both are linked above.

How often is this page updated?

When a study gets a new read or a new one publishes. The date at the top is the last change. Battles and page tests currently running on the site have reads booked for October and November 2026.

Sources

• SEOtesting archive readout, 8 August 2026: 130 tests, 108 finished with results (internal analysis, method and caveats summarised above).
• StudioHawk AI referral dataset, 1,203,748 sessions across 600+ businesses, 15 January to 14 July 2026; StudioHawk audit dataset, 1,000+ audits; StudioHawk 100-brand ecommerce GA4 dataset, July 2024 to March 2026.
• Bing AI Search report for lawrencehitches.com, 28 June 2026 (citation grounding map).
• Cloudflare verified-bot analytics, Search Console, Bing Webmaster Tools and SEOtesting for lawrencehitches.com, March to September 2026.