In July we published The Coverage Illusion, which showed that AI engines disagree about who belongs at the top of a category far more than anyone expected. That analysis was about answers: which brands the engines name and rank.
This one is about sources: the pages and domains engines actually cite when they compose those answers. We analyzed the citation corpus behind the August 2026 edition of the CiteHawk AI Index: 12,737 citation records across 192 categories and the 8 engines we track, plus search telemetry parsed from more than 12,000 collected answers across our monitored workspaces over the last 60 days. Every figure below carries its sample size, and every one is computed from our own corpus. Nothing here is an industry estimate.
The short version: the engines are not reading the same internet, most of what they cite is invisible to Google-first thinking, and the retrieval behavior behind AI answers is less predictable than the emerging folklore suggests.
Finding 1: Cross-engine citation overlap is 5% to 19%
For every pair of engines, we measured how much their cited domains overlap within the same category (Jaccard overlap: shared domains divided by all domains either engine cited), then averaged across the roughly 190 categories both engines covered.
The highest overlap between any two engines was 18.6% (Grok and Perplexity). The lowest was 5.1% (Gemini and ChatGPT). Every other pair fell between those numbers.
Flip it around and the picture is starker. We measured how much of each engine's citation diet is exclusive to it, meaning domains no other engine cited for the same category:
- ChatGPT: 67.4% of its cited domains were cited by no other engine (1,364 domain-category pairs)
- Gemini: 55.1% (2,142 pairs)
- Bing Copilot: 47.9% (1,732 pairs)
- Perplexity: 45.0% (3,052 pairs)
- Grok: 36.1% (1,404 pairs)
- Claude: 27.8% (976 pairs)
If your content earns citations on one engine, that says very little about whether the others will ever see it. Source strategy is per-engine work, and a tracker that watches one or two engines is measuring one or two reading lists out of eight.
Finding 2: Most AI citations live outside Google's top 10
For 38 sampled Index categories, we fetched Google's top 10 organic results for the exact category question each engine was asked, then checked how many of each engine's cited domains appeared in that top 10.
The overlap ranged from 13.1% (ChatGPT) to 29.2% (Grok), with Claude at 20.2%, Perplexity at 23.6%, and Gemini at 24.1%.
Which means 71% to 87% of the domains AI engines cite for a buying question do not appear on Google's first page for that same question. Ranking well on Google and being cited by AI are overlapping but substantially different games. If your visibility strategy starts and ends with organic rankings, most of the AI reading list is out of your frame.
Finding 3: The engines' community appetite differs by two orders of magnitude
We classified every cited domain in the Index corpus by source type using our deterministic classifier. The share of citations going to community sources (forums such as Reddit, Quora, and Stack Exchange) per engine:
- Gemini: 10.6% of 3,068 citations
- Perplexity: 7.4% of 3,795
- Grok: 3.1% of 3,263
- ChatGPT: 0.3% of 2,005
- Claude: 0.0% of 2,534
- Bing Copilot: 0.0% of 1,744
The practical consequence is that source recommendations should be conditional on which engine you are trying to win. A Reddit presence plausibly moves your visibility on Gemini, Perplexity, and Grok. On the August corpus, it does approximately nothing for Claude or Bing Copilot citations, because those engines essentially do not cite community sources for buying questions.
Finding 4: Fan-out queries are real, capturable, and not very deterministic
When an engine answers a buying question, it typically issues its own search queries first (the industry calls this query fan-out). Those queries are observable: across our monitored workspaces we captured 16,344 fan-out queries from the answers themselves, including 2,705 from Claude, 5,571 from Gemini, and 6,828 from Grok over 60 days.
Two things the data pushed back on:
Fan-out is not deterministic. Of 535 (engine, prompt) pairs we asked at least twice with search activity, only 16.3% produced the identical query set every time. The same engine, asked the same question, usually goes looking with different words. Optimizing for one observed fan-out query is optimizing for one roll of the dice.
Year-stamping is the exception, not the rule. The share of fan-out queries containing a year (2024 through 2029) was 26.0% for Claude, 12.5% for ChatGPT, 12.3% for Grok, and 3.3% for Gemini. Freshness matters to engines, but mostly they do not ask for it in the query string.
Finding 5: How often engines actually search
Share of answers where the engine issued at least one live web search, from our workspace corpus (roughly 1,400 to 1,560 answers per engine over 60 days):
- Perplexity: 100% (retrieval-native)
- Grok: 99.2%, averaging 5.23 searches per answer
- Gemini: 98.9%, averaging 3.96 searches per answer
- Claude: 91.6%
- ChatGPT: 79.6%
- DeepSeek: 0% (no live search at all)
One method note for honesty: our Claude configuration caps searches at one per answer, so Claude's search frequency is measurable but its search intensity is not comparable to Grok's or Gemini's on our corpus. And DeepSeek's zero is a structural fact worth knowing: anything it says about your brand comes from training data, not your current website.
Finding 6: Claude's citations align with Brave far more than with Google
An idea circulating in the industry is that Claude's web search leans on Brave's index. We tested the observable version of that claim on our corpus: for 60 sampled Index categories, we fetched Brave's top 10 results for the exact category question and checked how many of Claude's cited domains appeared there.
The answer: 49.5% of Claude-cited domains appear in Brave's top 10 for the same question, versus 20.2% in Google's top 10 (from Finding 2's sample). That is not the near-total alignment sometimes claimed, but it is roughly two and a half times the Google alignment, and it carries a practical implication most SEO programs ignore completely: if Claude visibility matters to you, your Brave ranking is a lever, and almost nobody is pulling it.
Finding 7: The reading list churns
Across our monitored workspaces, an average of 48.9% of the domains an engine cited in a given month were domains it had not cited for that workspace the month before (22 workspace-engine month pairs, ranging from 7.6% to 88.6%). The reading list is not a fixed target you earn your way onto once. It is a rolling audition.
This is also why we started archiving the Index's full citation corpus every edition: from September 2026 the Index will publish month-over-month citation churn at category level, computed the same way every month.
What this means if you own a brand
- Track every engine or accept blind spots. With 5% to 19% citation overlap and 28% to 67% exclusive sources per engine, single-engine measurement misses most of the picture.
- Do not assume Google rankings transfer. Most of what AI cites is not on Google's first page for the same question.
- Match the tactic to the engine. Community strategy for Gemini, Perplexity, and Grok. Answer-ready owned pages and authority sources for Claude and ChatGPT.
- Refresh matters even when the query does not say so. Engines re-choose their sources constantly; stale content loses auditions it never knows it attended.
Every number above is reproducible from the frozen Index editions, which publish their full answer corpora for independent recomputation. If you want to know which sources the engines cite in your category, and whether you are on that list, that is exactly what CiteHawk measures.
