Key findings
What we found
- On 54 of 100 searches the two engines cited no source in common at all. Across the whole set they cited 939 domains and shared only 106 — a Jaccard overlap of 0.113.
- Claude cited Reddit, YouTube and Quora exactly zero times. Not rarely — never, across 3,753 citations. Google cited Reddit for ~43 of the same searches and YouTube for ~68.
- 93% of Claude’s citations go to vendor blogs and listicles, against 58% for Google. Claude effectively has one format.
- One vendor’s programmatic listicle farm, guideflow.com, was cited for 38 of 100 searches — while ranking in Google’s organic top 30 for only 8 of them.
- Claude’s citations are far less anchored to Google’s rankings: 13.5% of what it cites ranks in Google’s organic top 10, against 38.5% of Google’s own AI Overview citations.
- Both engines are about equally fickle: ask again and roughly half the citations change. 50% of Claude’s citations repeated on all three asks, against 51% of Google’s.
What each engine cites, by source type
Share of distinct domains cited per answer, averaged over three asks. 542 distinct citations per ask for Claude, 614 for Google.
The same numbers, side by side
| Source type | Claude | Google AI Overview |
|---|---|---|
| Vendor blog or listicle Round-ups and “N best X” posts published by vendors. | 93% | 58% |
| Video YouTube. Claude cited no video at all. | 0% | 11% |
| Vendor product page A vendor’s own homepage or product page. | 1% | 11% |
| Forum or community Reddit and Quora. Claude cited neither, once. | 0% | 8% |
| News or media Mainstream and trade press. | 2% | 5% |
| Review aggregator G2, Capterra, TrustRadius and peers. | 3% | 3% |
| Analyst Gartner, Forrester, IDC. | 0% | 3% |
| Social or pro network LinkedIn, Medium and similar. | 0% | 1% |
| Professional, org or gov Trade bodies and government sites. | 0% | 1% |
The overlap
Two engines, two source sets
If AI visibility were one problem, the engines would be drawing on roughly the same pages. They are not. Winning a citation in one tells you almost nothing about the other.
54/100
Searches with zero shared sources
Neither engine cited a single page the other did. Jaccard overlap across the full set is 0.113.
437
Cited only by Claude
Against 396 cited only by Google, and just 106 cited by both.
5.4 vs 6.4
Distinct sources per answer
Claude cites 12.5 times per answer but repeats domains, landing on 5.4 unique.
The forum and video gap is absolute, not marginal. Across 3,753 Claude citations there is not one link to reddit.com, youtube.com or quora.com — no domain containing “reddit” or “youtu” appears anywhere in the dataset. Analysts are nearly as absent: gartner.com appears twice, for a single search. The playbook that earns Google AI Overview citations — seed the Reddit thread, publish the comparison video — does essentially nothing for Claude.
The outlier
One listicle farm won a third of Claude’s answers
guideflow.com sells demo-automation software. It also publishes one “N best X” post per software category — and that library, not its product, is what Claude keeps citing.
38/100
Searches it was cited for
From accounting software to infrastructure-as-code tools.
12.7
Answers per ask where it was cited first
More opening citations than any other domain in the study.
8/38
Of those, ranking on Google
It reaches Google’s organic top 30 for only 8 — positions 14–27, never the top 10.
39 URLs, one per category
| Cited path on guideflow.com |
|---|
| /blog/ab-testing-tools |
| /blog/accounting-software |
| /blog/age-verification-software |
| /blog/application-performance-monitoring-tools |
| /blog/best-business-intelligence-software |
| /blog/best-contract-lifecycle-management-software |
Titles follow a fixed template — “8 best age verification software for 2026”, “15 best A/B testing tools for SaaS in 2026”. Category-complete coverage, year-stamped.
Why
The queries behind each answer
Alongside every answer the API returns its fan-out queries — in DataForSEO’s words, “related search queries derived from the main query to provide a more comprehensive response”. They show the shape of query each answer was assembled against, and that shape lines up closely with the pages that win.
2.0
Fan-out queries per answer
Almost always exactly two: the keyword, then one reformulation.
51%
Queries with a year appended
Nearly always “2026”. Undated content is competing against dated content.
574
Queries captured
Across 300 answers, returned by the API alongside each response.
A sample of the fan-out queries
| Fan-out query |
|---|
| best ab testing tools 2026 |
| top A/B testing software comparison |
| best A/B testing tools 2026 |
| top AB testing software comparison |
| best ab testing tools 2026 |
| top A/B testing software comparison |
| best accounting software 2026 |
| best accounting software for small business |
| best accounting software 2026 |
| best accounting software for small business |
The shape is consistent: <keyword> 2026 plus top <category> comparison. Which is exactly the shape of the pages that win.
Full dataset
Explore the data
Every search with both engines’ sources side by side, every cited domain tagged by which engine found it, and the fan-out queries returned with each Claude answer.
| Claude cited | Google AI Overview cited |
|---|
No searches match these filters.
| Source type | Cited by |
|---|
No domains match these filters.
No queries match these filters.
Free content audit
You now need to win two different source sets
What earns a Google AI Overview citation is not what earns a Claude citation — the two engines share 0.113 of their sources. We’ll check what each one cites for the searches that matter to your product, and show you which gaps are realistically winnable.
How we ran it
Method and limitations
- The same 100 “best <software>” keywords were sent to both engines on 19 October 2026, three times each — 600 answers in total.
- Claude:
claude-sonnet-5with web search enabled, via DataForSEO’s/ai_optimization/claude/llm_responses/live. The prompt is the bare keyword, so no prompt design of ours leaks into the results. - Google: DataForSEO’s SERP API with
load_async_ai_overview=true, US, English, desktop. - Like-for-like rule. Claude has no citation panel — it cites inline, per sentence. So every comparison here counts distinct domains cited per answer, and counts Google across all of its citation surfaces rather than just the three visible in the panel. Using Google’s visible top 3 instead sharpens the contrast; we report the more conservative reading.
- Source types are assigned by the same rule-based labeller for both engines, so the comparison cannot be an artefact of different labelling. It agrees with a manually labelled set 97.6% of the time.
- Claude citations are the
annotationsattached to each sentence of the answer. Position beyond order of first appearance does not exist, so no ranking is implied. - Repeatability is measured at the level of a (search, domain) pair: of the citations an engine made, how many it repeated on all three asks.
- On fan-out queries. DataForSEO documents this field as “related search queries derived from the main query to provide a more comprehensive response”. We report them as what they are documented to be. We have not independently confirmed that they are the model’s verbatim web-search calls, so we do not claim that — though the ~14,000-token inputs on four-word prompts show substantial retrieved context behind each answer.
- Limitations: one model, one day, US-oriented bare-keyword prompts. Opus and Haiku may cite differently, and longer real-world prompts may too — both untested. 6 of 300 Claude answers cited nothing at all.
Companion study, same keyword set: what Google’s AI Overview cites for software searches.