Checked 30 October 2026
Nobody has published a controlled test of whether building backlinks gets a site cited more in AI answers. Every study I found is a correlation, and in the largest ones links come out weak: well behind how often a brand is mentioned across the web.
That doesn’t mean links don’t matter. It means the claim that they do, or don’t, rests on correlations between big brands and big numbers, measured mostly in the US, with no raw data anyone outside the vendor can check.
The short version
| Question | What the evidence supports |
|---|---|
| Do sites with more backlinks get mentioned more in AI answers? | Yes, weakly to moderately. Correlation only |
| Does building links cause more citations? | Unknown. No controlled study published |
| Do brand mentions matter more than links? | In Ahrefs’ 75,000-brand data, yes, on every engine tested |
| Do AI answers cite pages that rank in Google’s top 10? | Depends on the engine and the date. Anywhere from about 4% to 76% |
| Can I check any of this myself? | No. None of the 16 publishes raw citation data |
Links and authority versus being mentioned
| Study | Sample | What was compared | Finding (Spearman unless noted) |
|---|---|---|---|
| Ahrefs, Dec 2025, updated Aug 2026 | 75,000 brands | Brand mentions in ChatGPT, AI Mode and AI Overviews vs DR, backlinks, branded web mentions, branded search | Branded web mentions 0.66–0.71. Domain Rating 0.27–0.33. Backlinks described as very weak on every engine |
| Ahrefs, May 2025, updated Apr 2026 | 75,000 brands | Mentions in AI Overviews | Branded web mentions 0.664, referring domains 0.295, backlinks 0.218 |
| Semrush with Growth Memo, Oct 2025 | 1,000 domains | AI mentions across five AI surfaces vs Authority Score and links | Authority Score 0.57. Follow and nofollow links almost identical, about 0.50 |
| Seer Interactive, Mar 2026 | 62 brands, 4 industries | ChatGPT citations vs 37 SEO metrics | Domain authority 0.75 on average; total backlinks 0.69. Corrected for multiple testing; author flags brand size as a likely confounder |
| Seer Interactive, Jan 2025 | 10,000 questions | Brand mentions in GPT-4o API answers | Google page-one rank strongest (about 0.65). Backlinks weak or neutral |
| Rankability, Sep 2026 | 1,645 citations, 11 engines, Aug 2026 | Cited pages vs their links and rank | 72.8% of cited pages had no referring domains of their own. A quarter came from sites below DR 40 |
Two things stand out.
The big samples find links weak and the small ones find them strong. Seer’s 62 brands and Semrush’s 1,000 domains give 0.5 to 0.75; Ahrefs’ 75,000 brands give 0.2 to 0.3. Seer flags the likely reason itself: brand size. Bigger brands have more links and more mentions of everything. In a small sample of established brands, size drives both.
Follow and nofollow links correlate almost equally with AI mentions in Semrush’s data. If links were passing something to AI systems the way they pass PageRank, you’d expect a difference. A brand-size effect would produce exactly this result.
Rankings versus being cited
| Study | Engine | Share of AI citations from Google’s top 10 |
|---|---|---|
| Ahrefs, Jul 2025 | AI Overviews | 76% (1.9 million citations) |
| Ahrefs, Mar 2026 | AI Overviews | 38% (4 million URLs; parsing method changed from July 2025) |
| BrightEdge, Sep 2025 | AI Overviews | 54.5% overlap, rising from 32.3% over 16 months (sample size not given) |
| Semrush, Jul 2025 | AI Mode | About 35% of URLs, 54% of domains |
| Ahrefs, Aug 2025 | ChatGPT, Gemini, Copilot | About 12% on average |
| Chen et al., EDBT 2026 | GPT-4o, Gemini, Claude, Perplexity | 4% to 15% domain overlap |
| Seer Interactive, Feb 2025 | ChatGPT search | 87% match Bing’s top results; 56% match Google |
The spread is the finding. Google’s own AI features overlap with Google’s rankings far more than anyone else’s AI does, and even that figure halved between Ahrefs’ two studies. Ahrefs changed its parsing method in between, and Google moved AI Overviews to a new model in January 2026, so the two can’t be separated. ChatGPT lines up with Bing, which makes sense if it searches Bing. “Rank well and you’ll be cited” is roughly true for AI Overviews, sometimes, and much less true elsewhere.
Scored on five checks
I scored all 16 on the same five checks I used for the 32 most-repeated AI search statistics: does the study say when it measured, how many, where, how (in full), and is the raw data published?
| Check | Studies passing, of 16 |
|---|---|
| When (measurement period) | 5 |
| How many (sample size) | 15 |
| Where (country) | 1 |
| How (method in full) | 4 |
| Raw data published | 0 |
Only one study names the country of its sample: Zhang et al. (December 2025), using US trending queries, which tested domain popularity rather than links. None publishes the raw citations. Rankability publishes aggregate CSVs; Zhang et al. offer data for research use on request.
Not one of these says what the UK looks like.
The two closest things to an experiment
- A Stacker and Scrunch test (December 2025) placed the same eight stories on a brand’s own site and syndicated them to other publishers. The citation rate rose from 8% to 34%. Syndication changes reach, mentions and links at once, the engines aren’t named, and Stacker sells syndication.
- A link-building vendor’s case study (December 2025) on one site has no control group and changed content at the same time.
Neither isolates links.
What I’d do with this
- Treat “links drive AI citations” and “links don’t matter for AI” as unproven. Both are being sold.
- Brand mentions across the web are the most consistent correlate in the largest dataset. They’re also the hardest of these to manufacture.
- If you’re optimising for AI Overviews specifically, ranking still overlaps more than it does anywhere else. For ChatGPT, look at Bing.
- Ask any vendor quoting one of these numbers the five questions above. If the answer to “where” is blank, it isn’t a UK number.
What a real test would look like
This is a method, not a result. Nothing below has been run.
- Take 40 pages on comparable sites in one topic, none currently cited for a fixed list of 20 UK questions.
- Split them at random: 20 get new links from relevant sites over eight weeks, 20 get nothing. No content changes to either group.
- Run the 20 questions twice a week on four engines, signed out, from the UK, before and after.
- Compare the change in citation rate between the two groups. Publish every answer.
It’s expensive, slow and small. It would still be the first public evidence of cause rather than correlation. If you run something like it, send it to me and I’ll link it.
Limits
- This is a reading of published studies. I haven’t rerun any of them.
- Most were published by companies that sell SEO tools, link building or AI visibility tracking.
- Correlation coefficients from different studies use different samples, metrics and engines. Putting them side by side shows a range, not a single answer.
- SE Ranking published a widely quoted ChatGPT ranking-factors study putting referring domains at the top. Its page didn’t open for me, so it isn’t scored here.
- The scores are reading judgements. Each study’s scores and a note are in the CSV.
Download
CSV: 16 studies, what each measured, and its five-check scores. One row per study. Free to cite with a link.
Related: AI search statistics, traced to the source · Crawled is not cited · Five labels every AI visibility number needs · What do AI visibility trackers say about how they measure?
Sources
Opened on 30 October 2026. Each study is linked in the tables above. Also: