What 34 Studies Reveal About AI Search in 2026
77 studies scored. 34 made the cut. Here's what they say. (~15 min read)
Over the past few months, AI search has been moving faster than most can keep up. Like a lot of you, I’ve been hoarding tabs: saving studies, bookmarking data, collecting without actually taking the time to read and think through it all. I’m that person with 2,852 saved LinkedIn articles and 100+ tabs opened at any given time.
One slow weekend, I decided to sit down properly and read through some of them. One study became ten, ten became twenty, and somewhere around study thirty I realised I had something worth sharing.
There’s no shortage of AI search studies right now. The problem is quality. Vendor research with convenient conclusions, small samples dressed up as signals, methodology buried or missing entirely. More research also means more noise.
This piece is for brand leadership, SEO leads, and digital strategists who’ve been trying to keep up with AI Search but don't have time to read everything.
To my knowledge, nothing has been done at this scale: 77 studies scored against a weighted methodology, 34 kept and analysed.
If you’ve been struggling to get a clear picture of where AI search actually stands in 2026, this is the one piece you need to read.
Hi! I’m JY. This is my first Substack post.
I’ve spent close to 15 years in SEO and led one of the larger SEO teams in Australia, working with some of the biggest brands in the country. I recently went independent, which means I now have more time to read studies (and fewer excuses not to).
Subscribe to hear my irregular takes on SEO and AI Search!
Given everything I said about poor research, I figured I should show my work first.
How 77 studies became 34
I started with studies I’d already saved, then went looking for more with the help of my colleague Claude. Landed on a total of 77 published between January 1st and March 31st 2026. Each was scored against five weighted criteria:
34 made the cut. 43 didn’t. Most of what got cut failed on transparency or sample quality, not necessarily size. Some smaller studies with clear methodology outscored larger ones that didn’t show their working.
List of brands cited in this article: a16z, Ahrefs, Amsive, AirOps, BrightEdge, Chartbeat, Conductor, Dan Petrovic (DEJAN), Define Media Group, Gauge, Graphite, Gumshoe.ai, Harvard/NBER, IBM, Kevin Indig (Growth Memo), NRF, OtterlyAI, Peec AI, Previsible, Profound, Promptwatch, Reuters Institute, SE Ranking, Seer Interactive, Semrush, Similarweb, SparkToro, Superlines, Tinuiti, Wix
You can find the full list of sources at the end of this article.
If you only have two minutes, here’s the TL;DR
Search traffic declined 2.5%, not 25%. Total discovery, search + AI combined, grew 26%. Search got bigger.
Position 1 CTR dropped 58% when an AI Overview appears. Clicks redistributed within the page, not off it. Branded queries got a boost. Users aren’t abandoning search.
Only 38% of AI citations come from top 10 ranked pages. Ranking well no longer means AI will cite you. A page ranked 19th can outperform a page ranked 1st.
50% of consumers have purchased after AI research. AI starts the purchase journey. It rarely finishes it. The channel earns the assist, not the conversion.
Only 4% of AI citations come from a brand's own website. The rest comes from its ecosystem, in particular YouTube, Reddit, and third-party review sites. When a brand isn’t already in the model’s training data, citation rate drops 5x.
When AI recommends a brand, 68% of consumers verify by Googling it. That click shows as organic search in your analytics, not AI. The influence is real. The attribution is invisible.
Here’s what I’m covering below:
The search market is bigger than you think
Search traffic declined 2.5%, not 25%
83% of AI usage happens in apps
Total discovery grew 26%
Position 1 lost its edge, but not every click disappeared
CTR loss is real
Branded queries tell a different story
The variance by vertical is massive
AI visibility runs on different mechanics than search
Pattern1: The Citation Divorce
Pattern 2: The Ghost Citation
Pattern 3: The retrieval Gap
Pattern 4: The Chunking Effect
Pattern 5: The Ski Ramp
What this means for your content
Where AI looks first
What kind of content gets selected
What the evidence doesn’t support
The purchase journey is now AI > verify > buy
AI is already part of the journey
AI extended the research process
The Verification Loop
The attribution problem
Fragmented market, loyal users
AI isn’t just search
The market is fragmenting
But most users are not platform-hopping
Disclaimer:
This is a review, not original research. I didn’t run any of these studies. I’m synthesising other people’s work, with all the limitations that come with that. No single stat tells the full story.
I used Claude as a research and fact-checking sidekick throughout this. The thinking, writing and conclusions are all mine.
Theme 1: The search market is bigger than you think
Let’s start with the stat that started all the panic.
Search traffic declined 2.5%, not 25%
Gartner predicted in 2023 that organic search traffic would fall 25% by 2026 due to AI. Based on the data we have from Q1 2026, that prediction hasn’t landed (yet).
In January 2026, Graphite and Similarweb published what I’d consider the most rigorous study on this question: 40,000+ of the largest US websites, validated against first-party Google Search Console data with a 0.86 median correlation. That’s a high number, meaning the dataset tracks reality closely rather than extrapolating from a small sample.
The result: SEO traffic declined 2.5% year-over-year. Google traffic specifically was up 0.8%. The ten largest sites grew 1.6%. The decline is concentrated in mid-tier sites, roughly ranked 100 to 10,000, the long tail of informational content that AI Overviews now cover.
Most of the alarming traffic numbers you see online come from surveys with small samples, vendor studies with obvious incentives, or real but unrepresentative experiences reported as industry wide trends.
Publishers are a different story. Reuters Institute and Chartbeat tracked 2,500+ publisher sites and found Google search traffic to publishers dropped 33% globally. Small publishers got hit more than smaller ones.
83% of AI usage happens in apps
Graphite’s own January study measured web traffic only. When they added mobile app data in their March follow-up, the number jumped from roughly 10% to 56% of search volume. The difference: 83% of AI usage happens in apps.
AI now receives 45 billion monthly sessions worldwide. That’s 56% the size of search globally, 34% in the US. If your measurement framework only tracks web referrals, you’re seeing about a fifth of actual usage.
Total discovery grew 26%
Combine search engine sessions with search-related AI prompts and total discovery activity grew 26% worldwide between Q1 2023 and Q4 2025, and 16% in the US (Graphite, 2026).
Search didn’t shrink. AI just added more on top.
Using AI and clicking through from AI are two completely different things. Ahrefs tracked 76,000 websites and found ChatGPT handles roughly 12% of Google’s search volume but sends 190 times less traffic to websites, with a CTR 96% lower than Google’s.
People are using AI and traditional search more than ever. They’re just not clicking through as much.
Theme 2: Position 1 lost its edge, but not every click disappeared
AI Overview coverage fluctuated significantly through 2025, settling at around 16% of queries by year end (Semrush, 2026). By Q1 2026, the picture had stabilised enough to measure properly.
CTR loss is real
Ahrefs tracked 300,000 keywords over two years. On queries where an AI Overview now appears, position 1 CTR dropped from roughly 7% to 1.6%. But CTR was already falling before AIOs arrived. Ahrefs separated out the AIO-specific effect and put it at around 58% of the total decline. Other positions are also affected, but not as much as position 1.
Seer Interactive ran a separate analysis across 42 organisations and 25.1 million impressions and found similar impact: organic CTR down 61% and paid CTR down 68% on AIO-affected queries.
But Seer also found organic CTR fell 41% on queries with no AI Overview at all.
AI Overviews seem to have accelerated a trend instead of starting it.
Branded queries tell a different story
Branded queries tend to benefit from AI Overviews. Amsive found an +18.68% CTR lift last year, and Seer discovered that brands cited in an AI Overview saw +35% organic and +91% paid CTR compared to uncited brands on the same queries.
The reason is simple: for informational queries, AI answers the question and removes the need to click. For branded queries, the user already knew where they were going. AI just accelerates that choice.
If AI already knows your brand, you get the click. If it doesn’t, someone else does.
The variance by vertical is massive
Exposure and referral traffic are split sharply by industry, and the two don’t move together.
Healthcare triggers AI Overviews on 48.75% of queries, the highest of any vertical in the studies reviewed here, but generates only 0.63% AI referral traffic, the lowest. Maximum exposure, minimum clicks. AI tends to answer health queries completely without needing to send anyone anywhere.
Technology sits at the opposite end of that relationship: lower AIO exposure at around 15% of queries, but the highest AI referral traffic of any vertical at 2.8%. High exposure and high traffic do not move together (Conductor, 2026).
Real estate has the lowest AIO trigger rate at 4.48%. Finance lands at 25.79%. Both reflect the same pattern: queries that are local, transactional, or high-stakes tend to resist AI summary for now (Conductor, 2026).
AI Overviews are also expanding beyond informational answers. Commercial query coverage grew from 8.15% to 18.57%, navigational from 0.74% to 10.33% (Semrush). The feature started where it could answer safely. It is now moving toward the queries that drive revenue.
E-commerce is already feeling it. AI Mode shopping triggers on 61.7% of product searches, with eBay capturing 86.6% of AI Mode product listings. Marketplace presence is becoming an AI visibility factor in a way that brand websites are not (SE Ranking, 2026).
Theme 3: AI visibility runs on different mechanics than search
This section is the most research-heavy part of the piece. Key patterns emerge from the data that improve our understanding of how brands get surfaced in AI platforms. Some came with names attached. Others didn’t, so I gave them one. Patterns are easier to act on when you can actually refer to them but I’m open to better suggestions!
Pattern1: The Citation Divorce
The overlap between organic top-10 rankings and AI citations collapsed from 76% to 38% in seven months (Ahrefs, March 2026). Ranking highly no longer guarantees AI will cite you. The two systems are increasingly running on different logic.
Ahrefs tracked 863,000 keywords and found that a third of AI citations now come from positions 11 to 100. Another third from outside the top 100 entirely.
Why? Google rankings are essentially a popularity contest. The more credible sites that link to yours, the higher you rank. AI works differently: it prioritises the most relevant passages, not the most authoritative pages. A page at position 47 with a clearly structured answer is more likely to be cited than a position 1 page that buries the same answer in paragraph twelve.
You can rank 1st and be invisible to AI. You can rank 40th and be cited constantly.
The implication: rankings and citations now measure different things. SEO foundations still matter for both, but ranking well no longer means you’re visible.
Pattern 2: The Ghost Citation
Most people assume citations drive recommendations. The data suggests it works the other way.
Seer Interactive ran six behavioural tests across 541,213 LLM responses across 20 brands and 6 platforms, and found that when a brand appeared in a recommendation, it was also cited 53.1% of the time. When the model chose not to recommend that brand, the citation rate dropped to 10.6%. A 5x gap driven by whether the model decided to name you, not whether your content was there.
The theory: the model decides who to recommend first, based on what it already knows from training, then finds sources to support that. LLMs are probability machines. Recommendation and citation are generated at the same time, not one after the other.
Britney Muller (actual AI expert) described this as a form of “post-hoc rationalisation” in a discussion with Rand Fishkin on LI. The model generates an answer based on learned patterns, then produces citations alongside it rather than deriving the answer from them. Rand later expanded on how LLMs generate responses.
SparkToro and Gumshoe found the same pattern from a different angle. Across 2,961 prompt runs, the same brands kept showing up in 55 to 77% of responses, no matter how the question was phrased. AI keeps going back to the same names it already knows and trusts.
If you’re not in the training data, you’re rarely cited. If you are, you show up everywhere.
Pattern 3: The Retrieval Gap
AirOps analysed 548,534 pages retrieved by ChatGPT across 15,000 prompts and found that 85% never appear in the final response. The model sees far more than it shows.
Getting retrieved and getting mentioned are two different things.
As the model processes a query, it breaks it into multiple hidden sub-questions. AirOps found that 33% of citations come from these fan-out queries, not the original prompt.
95% of those sub-questions have zero search volume. They don’t show up in keyword tools, but they influence what gets cited.
A brand might appear not because it matches the user’s query, but because it wins in a question the user never asked.
Pattern 4: The Chunking Effect
AI models don’t read pages. They read fragments.
Dan Petrovic at DEJAN reverse-engineered Google’s grounding pipeline (across 7,060 queries and 2,275 pages but client data so not independently replicated) and found the grounding budget is fixed at around 2,000 words per query, split across cited sources by rank. The #1 ranked source gets 531 words extracted; the #5 source gets 266. For pages over 3,000 words, only 13% of content makes it through.
Most of your content is never seen as a whole. Pages are broken into small chunks and evaluated independently. What gets used is not the page, but the passages that survive extraction.
The practical implication is simple: each section of your content needs to make sense in isolation. A paragraph that relies on context from three sections earlier will be extracted without that context and may be skipped entirely. Write as if every block could be the only block the AI reads.
Pattern 5: The Ski Ramp
The structure AI rewards is simple: lead with the finding, follow with the evidence.
In an analysis of 1.2 million ChatGPT responses and 30 million citations, Kevin Indig described what he calls the “Ski Ramp” effect: 44.2% of citations come from the first 30% of a page’s content
ChatGPT was predominantly trained on journalism and academic papers, both of which follow BLUF: Bottom Line Up Front. It learned that the most important information lives at the top and reads accordingly.
But within a section, it reads deeply: 53% of citations come from the middle of a paragraph, not the first sentence. The model reads carefully, the problem is what it never reaches.
If your key claim is buried in paragraph eight, it’s invisible to AI (and probably users too).
What this means for your content
1. Where AI actually looks
AI starts with your ecosystem, not your website. Across 27 million citations, around 4% of AI citations come from a brand’s own website. The other 96% comes from third-party sources AI already trusts (Profound, 2026).
Promptwatch tracks real prompt responses across both Google AI Mode and AI Overviews every month. Across both surfaces, the pattern is consistent for Q1: YouTube and Reddit dominate the citation table, followed by LinkedIn, Facebook, and Google itself.
This is consistent with what multiple studies show independently. AI overwhelmingly pulls from UGC platforms and high-authority aggregators over brand-owned content.
YouTube: 40% of cited videos have fewer than 1,000 views and subscriber count barely matters (OtterlyAI). What drives 94% of citations is long-form, transcript-rich content. Depth and clarity beat reach, especially given YouTube transcripts are part of model training data.
Reddit: Citation share grew 73% from Oct 2025 to Jan 2026, making up 44% of AI Overview social citations (Tinuiti). 99% point to specific threads, not brand pages. AI is pulling answers from conversations, not platforms.
LinkedIn: Frequently cited for B2B and expert-led queries, with individual posts surfacing more than company pages. Authority comes from perceived expertise and clarity of opinion, not traditional domain strength.
Reviews actively shape which brands get trusted and recommended. Around 34.5% of AI Overviews cite at least one review platform, and a small group of platforms dominate that visibility, with players like G2 consistently among the most cited sources (SE Ranking).
The implication: visibility is ecosystem-driven. If you’re not present across these platforms, you’re not in the consideration set.
2. What kind of content gets selected
Format determines whether you get cited.
Wix and Peec AI analysed 75,000 AI answers and 1.06 million citations (2026). Listicles earned 21.9% of citations, general articles 16.7%, and product pages 13.7%, totalling 52% of all citations across those three content types.
For commercial queries, listicles took 41% of citations, and in professional services, 81% of those came from third-party sites rather than brand-owned content.
The format that gets cited depends on what the user is trying to do. Match format to intent.
Listicles take the largest share at 21.9%. Blog content overall accounts for more citations than every other content type. The blog isn't dead. (just don't go creating hundreds of AI-generated listicles off the back of this).
3. What the evidence doesn’t support
SE Ranking analysed 300,000 domains and found no correlation between LLM.txt presence and citation frequency. OtterlyAI ran a 90-day experiment and found the file was requested in just 0.1% of crawler visits.
This aligns with Google’s current stance. LLM.txt is not a recognised or supported standard, and there’s no evidence it influences how content is selected.
Not every “AI SEO/GEO” tactic matters.
Theme 4: The purchase journey is now AI > verify > buy
Understanding how AI selects content is only half the problem. The other half is what happens after a user sees your brand.
AI is already part of the buying process
45% of consumers use AI during buying journeys (IBM and NRF, 18,000 consumers across 23 countries, January 2026)
50% have made a purchase after AI research (Semrush, 1,030 US shoppers, March 2026)
AI is not experimental anymore and is already influencing consideration.
AI extended the research process
What the traffic numbers don’t show:
77% use AI and traditional search together, not one replacing the other (Semrush, 2026).
87% say AI summaries help them understand brands faster (Semrush, 2026).
People aren’t switching from search to AI. They’re doing both, and doing more research overall. The “AI killed Google” narrative assumes a zero-sum game. The data suggests the opposite.
Visibility is being stretched across more touchpoints, not fewer.
The Verification Loop
86% of consumers verify AI recommendations elsewhere before acting. An AI mention is the start of consideration, not the end.
Where they go next:
68% check Google
48% visit the brand’s website
35% check review sites
35% check YouTube
AI introduces and other channels validate.
The attribution problem
That verification loop is creating an attribution black hole. Only 22% of AI-influenced purchases happen directly inside an AI tool (Semrush, 2026). The rest flow through channels that look like organic search, direct traffic, or referral.
When a user ask ChatGPT for a recommendation:
Your brand appears
They Google you: organic search
They visit your site: direct traffic
They check G2/Trustpilot: referral
At no point does “AI” appear in your analytics. The influence is real, but the trail is invisible.
Theme 5: Fragmented market, loyal users
AI isn’t just search
Most AI usage has nothing to do with search.
A OpenAI/Harvard study analysing 1.5 million ChatGPT messages found:
49% are “Asking” (questions, research, decisions)
40% are “Doing” (emails, code, summaries)
11% are “Expressing” (writing, brainstorming)
Only the Asking category overlaps meaningfully with traditional search behaviour, and that’s less than half of total usage.
The market is fragmenting
ChatGPT’s traffic share dropped from roughly 75% in early 2025 to around 60% by 2026. Gemini went from 5% to more than 20% over the same period. Copilot, Perplexity, Claude, and DeepSeek are each growing, each with distinct user bases (Similarweb, March 2026).
One caveat: Similarweb measures web traffic. As covered in Theme 1, 83% of AI usage happens in mobile apps. The real picture, across mobile and embedded tools like Microsoft 365, is almost certainly more fragmented than the chart shows.
But most users stay with one platform
Only around 20% of ChatGPT users also used Gemini in the same week. The majority stay within one tool (a16z, March 2026).
Switching costs are real: conversation history, personalisation, workflow integration. Anyone who has tried moving three years of ChatGPT context to a new tool knows the pain.
Your audience has probably already picked their platform. The work is showing up in the right one.
What this all means
Thirty-four studies don’t resolve every question around AI search, but they they make the picture a lot clearer.
The search is dying narrative doesn’t hold up. Total discovery grew 26%. People are using both systems, doing more research overall, and clicking less.
The key finding is that AI search runs on different logic. Citation Divorce, Ghost Citation, the Ski Ramp: these aren’t SEO concepts with new names. They describe a system where rankings and AI presence have decoupled, where your website is barely considered, and where the brands showing up consistently are already embedded across the sources the model draws from.
Two systems are now running in parallel. Understanding how each one works is becoming a requirement if you want to maintain visibility and influence in both. For some teams, that means modernising an existing SEO strategy. For others, it means building one.
No one has this fully figured out. The studies in this piece are a starting point, not a playbook. Testing and learning in your own context is still the best way forward.
I'll be back in a few months to see how much has changed.
FAQ
How much has search traffic actually declined? 2.5% year-over-year, not the 25% Gartner predicted. Publishers are down 33%, which is where the scary numbers might come from. Total discovery (search + AI combined) grew 26%.
Does our website even matter anymore? It matters, but it's not the only place AI looks. Your website accounts for 4.3% of AI citations. YouTube, Reddit, and third-party review sites dominate. Your website supports conversion but your ecosystem drives AI visibility.
Should we still be investing in SEO? Yes. Around 80% of what earns AI citations overlaps with solid SEO practice: authority, structure, technical fundamentals. But only 38% of AI citations come from top 10 rankings. Traditional SEO alone caps how much AI visibility a brand can achieve. A strong integrated team may not need a completely separate strategy, but it does need dedicated headspace to experiment with that 20%.
Is AI actually influencing purchasing decisions? Yes. 50% of consumers have made a purchase after AI research, and 45% use AI somewhere in the buying process. But 86% verify before buying, so AI earns the introduction, not the conversion.
How would we know if AI is affecting our business? You mostly wouldn’t, with standard analytics. AI-influenced journeys show up as organic, direct, or referral traffic. A metric worth tracking is AI topical presence: Are you consistently linked to the key commercial topics in your category? Because that tells you whether you’re actually part of the conversation or not.
What’s the biggest mistake companies are making right now? Measuring AI impact through referral traffic and concluding it doesn’t matter. AI referral traffic is around 1% of web traffic, but 50% of consumers have purchased after AI research. That gap is where the influence lives.
Sources
1. Graphite + Similarweb, “Debunking The Myth That Search Is Dying,” January 2026. 40,000+ US websites, GSC-validated (0.86 correlation).
2. Graphite, “AI Is Much Bigger Than You Think,” March 2026. Web and mobile app data, Similarweb panel.
3. Ahrefs, “AI Overviews Reduce Clicks by 58%,” February 2026. 300,000 keywords, GSC data.
4. Seer Interactive, “AIO Impact on Google CTR: December 2025 Update.” 42 organisations, 25.1M impressions.
5. Seer Interactive, “Ghost Citation Analysis,” 2026. 541,213 LLM responses, 20 brands, 6 AI platforms.
6. Ahrefs, “AI Overview Citations Drop From 76% to 38%,” March 2026. 863,000 keywords, 4M AIO URLs.
7. IBM + NRF, “AI and Consumer Purchase Journey,” January 2026. 18,000 consumers, 23 countries.
8. Semrush, “AI Tools and the Modern Buyer Journey,” March 2026. 1,030 US shoppers surveyed.
9. a16z, “State of Consumer AI 2025,” December 2025. Yipit data, subscription and retention analysis. [Q4 2025]
10. Growth Memo (Kevin Indig), “The Science of How AI Pays Attention,” February 2026. 1.2M ChatGPT responses, 30M citations, Gauge partnership.
11. Reuters Institute + Chartbeat, “Journalism and Technology Trends 2026,” January 2026. 2,500+ publisher sites.
12. Similarweb, “Gen AI Stats 2026 / Brand Visibility Index,” March 2026.
13. OpenAI + Harvard (NBER), “How People Use ChatGPT,” December 2025. Approximately 1 million de-identified messages. [Q4 2025]
14. Conductor, “2026 AEO/GEO Benchmarks Report,” January 2026. 13,770 domains, 3.3B sessions, 100M+ citations.
15. Semrush, “State of Search 2026,” Various 2025–2026. 10M+ keywords, Datos clickstream data.
16. Amsive, “Branded Query AIO Impact,” 2025. 700,000 keywords.
17. SE Ranking, “Google AI Mode Self-Citation: 17%,” February–March 2026. 68,313 keywords, 1,321,398 citations.
18. SE Ranking, “AI Citation Factors: ChatGPT and AI Mode,” November–December 2025. 2.3M pages, 295K domains. [Q4 2025]
19. Ahrefs, “AI Assistants Prefer to Cite Fresher Content,” February–March 2026. 17M citations across 7 platforms.
20. Ahrefs, “Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews,” February–March 2026. 75,000 brands analysed.
21. OtterlyAI, “YouTube Citation Study 2026,” March 2026. 100M citation instances, 6 platforms.
22. OtterlyAI, “The AI Citation Economy Report,” February 2026. 1M+ citations, ChatGPT, Perplexity, Google AIO.
23. AirOps, “The Influence of Retrieval, Fan-out, and Google SERPs on ChatGPT Citations,” March 2026. 548,534 retrieved pages across 15,000 original prompts and 43,233 total queries.
24. Previsible, “2025 State of AI Discovery Report,” February 2026. 1,963,544 LLM sessions, 9 industries, November 2024–November 2025.
25. SparkToro + Similarweb, “AI Tools as a Fraction of Web Traffic,” 2026. Datos clickstream analysis.
26. Wix + Peec AI, “Content Types in AI Citations,” 2026. 75,000 AI answers, 1.06M citations.
27. SparkToro + Gumshoe.ai, “AI Brand Recommendation Inconsistency,” January 2026. 2,961 prompts, 600 volunteers, 12 product categories.
28. BrightEdge, “AI Overviews at the One-Year Mark,” February 2026. Twelve-month analysis (February 2025 to February 2026) using BrightEdge Generative Parser. Industry-specific query tracking across nine verticals.
29. Tinuiti + Profound, “Q1 2026 AI Citation Trends Report,” March 2026. Nine commercial categories, seven AI platforms, October 2025–January 2026.
30. Superlines, “Citation Variance Across Platforms,” March 2026. 62 brands tracked across 10 platforms.
31. Define Media Group, “Google Discover Traffic Analysis,” March 2026. Publisher analytics data.
32. DEJAN (Dan Petrovic), “Google’s Grounding Pipeline Reverse-Engineered,” 2026.
33. Promptwatch, “Google AI Overview Citation Share,” January 2026. 568,499 prompts, 5,585,499 citations. Follow-up study covering AI Mode, February 2026.
34. Profound, “Who Shapes AI Answers? Enhanced Citation Categories,” January 2026. 27 million citations across ChatGPT, Gemini, and AI Overviews.
Methodology
77 studies scored on five criteria: sample size (x1.2), methodology transparency (x1.2), independence & bias (x1.0), actionability (x0.8), and industry discussion (x0.8). Max score: 25. 34 studies made the cut. A handful of studies from Q4 2025 are included where no more recent equivalent was available and are flagged in the sources section above. All other studies are from 1st January to 31st March 2026.
good stuff, i appreciate the weighting in your methodology