AEO and listening
Sources cited in AI answers about developer tools
Large citation studies combine many topics and industries. Their domain rankings are useful benchmarks, but they do not identify the sources cited for your developer-tool category. Build that list from a defined prompt sample.
Takeaways
- Compare citation studies only after checking their reporting periods and denominators. Share among the ten most-cited domains differs from share of all citations.[1]
- Choose channels from your buyers' questions. Collect comparison, implementation and troubleshooting prompts from support tickets, GitHub issues and community threads, then record which sources each AI engine cites. Broad domain rankings do not tell you where your developer-tool category gets its answers.[2]
- Use the citation map to choose your next task. Read the frequently cited pages and threads for your prompts. Fix missing or unclear documentation, or answer a relevant open question with a working example. Prioritize helping the reader, not placing your product name wherever citations appear.
- Make technical answers useful enough to reuse. Add runnable examples, clear definitions and fair comparisons where they help solve the question. A measurement study found more source content carried into answers from pages with code, statistics, definitions and comparisons, but it did not prove that adding those elements causes more citations.[5]
- Check patterns, not one answer. Repeat the same prompts on each engine and record the search mode, date and share of runs citing each domain. Sources vary across runs and reporting periods, so one result is a weak basis for changing your channel strategy.[3][9]
- Keep citation visibility separate from business results. Track whether answers cite your pages or name your product, then measure visits and purchases separately. A citation shows that a source was used; it does not establish product consideration or a sale.
Which websites do AI search engines cite most?
The studies below report different citation distributions across ChatGPT, Perplexity and Google AI Overviews. Their reporting periods and denominators are important to the comparison.
Profound analyzed 680 million citations from ChatGPT, Google AI Overviews and Perplexity between August 2024 and June 2025.[1] Wikipedia led ChatGPT with 7.8% of all citations and 47.9% of the share among its ten most-cited sources. Reddit led Perplexity with 6.6% of citations and 46.7% of top-ten share. Google AI Overviews spread its citations more evenly, with Reddit at 2.2% and YouTube at 1.9%.
Peec counted 30 million sources cited directly in answers across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews in the US.[2] Its March 2026 top ten reads Reddit, YouTube, LinkedIn, Wikipedia, Forbes, G2, Yelp, Facebook, Medium and TechRadar.
Report citation share separately by engine. The chart uses share among each engine's ten most-cited domains, which is different from share of all citations.
Why do the citation lists keep changing?
Citation distributions can change between reporting periods. A shift in the observed domain mix does not, by itself, identify the retrieval or ranking change that caused it.
Semrush tracked 230,000 prompts across ChatGPT search, Google AI Mode and Perplexity from July to October 2025.[3] In early August, Reddit appeared in close to 60% of ChatGPT responses. By mid-September it was around 10%. Wikipedia fell from roughly 55% to under 20% in the same window. Semrush did not confirm a cause. One theory points to Google removing a results parameter that third-party tools relied on. Semrush's own analyst suspected ChatGPT was trying to avoid over-citing a few sites.
ChatGPT responses citing Reddit: August to September 2025
Licensing agreements are relevant background, but do not establish why a domain receives a particular citation share. Columbia Journalism Review reports data licensing partnerships between Reddit and both Google and OpenAI.[4] A partnership is not evidence that a particular Reddit thread was retrieved for a particular answer.
Repeat the same prompt sample over time. Record the engine, search mode and collection date so a change in citation share can be investigated rather than treated as a permanent ranking factor.
How many sources does an AI answer actually use?
The number of citations and the amount of source material used in an answer are different measurements.
A 2026 measurement study ran 602 controlled prompts across ChatGPT, Google AI Overviews and Perplexity and collected 21,143 citations.[5] Mean citation counts were 16.35 for Perplexity, 12.06 for Google AI Overview/Gemini and 6.88 for ChatGPT. The cleaned sample contained 602 prompts for each of the first two platforms and 587 for ChatGPT. ChatGPT cited the fewest pages but each cited page shaped more of the final answer. The authors call this absorption: how much of a cited page's language, facts and structure ends up in the response.
Within that study, pages containing code snippets, statistics, definitions and comparisons had higher measured citation absorption. Pages containing code had a 76.88% higher mean influence score than pages without it; the corresponding difference for numbers or statistics was 61.55%. Question-and-answer (Q&A) formatting alone had a 5.74% lower mean score. These are descriptive comparisons, not effects established by randomized edits.
Citation absorption is the study's measure of how source content influences an answer; it is not the same as a direct quotation, a visit or a conversion. Treat the reported uplifts as findings from this experiment, not expected gains for every documentation page.
Correlation between brand mentions and AI visibility
In Google AI Overviews, mentions correlate more strongly with visibility than links do.
Ahrefs selected 75,000 brands using domains with Domain Rating above 40 and a qualifying keyword with at least 800 monthly searches. It excluded the roughly 26% with no AI Overview mentions from the correlation analysis.[6] Branded web mentions had the strongest correlation at 0.664. Branded anchors came next at 0.527. Domain Rating scored 0.326 and raw backlink count scored 0.218. Ahrefs is clear that these are correlations, not proof of cause.
Correlation with AI Overview brand visibility
These correlations support measuring brand mentions alongside citations. They do not show that posting another brand mention will improve visibility, or that any particular thread will be retrieved.
Limits of applying general studies to developer tools
Technical evaluations can involve version-specific documentation, issue trackers and implementation examples. The broad studies cited here do not isolate those prompts, so their overall domain rankings should not be presented as developer-tool rankings.
Three things set developer tool prompts apart.
- Comparisons are one useful prompt category. Developers ask "X vs Y" and "best Z for Y". Among the informational queries in Seer's 53-brand dataset, AI Overviews appeared for 267 of 280 comparison queries (95.4%) and 1,214 of 1,413 question-format queries (85.9%).[7]
- The sources are specialized. GitHub issues, official docs, Stack Overflow threads and engineering blogs carry the evidence. Few of those rank in a general top ten.
- The audience checks the work. In the 2025 Stack Overflow survey, 46% of developers said they distrust the accuracy of AI tools, against 33% who trust it.[8] That finding supports checking technical claims. It does not measure click-through behavior or compare the commercial value of citations across industries.
Peec makes the same point from the other side. Its general ranking is a starting point, and industries have their own dominant sources.[2] The cited studies do not supply a developer-tool source list. A category-specific prompt sample can provide one, with its own coverage limits.
How do you map the sources AI cites in your category?
Run a fixed set of real prompts across the engines on a schedule, log every cited URL, and group the domains by source type.
Write the prompt set
Pick 40 to 100 prompts your buyers actually ask. Use four shapes so you cover the whole decision.
- Best for: "best observability tool for a small Kubernetes team"
- Versus: "Datadog vs Grafana Cloud for startups"
- How to: "how to add rate limiting to a Next.js application programming interface (API)"
- Problem: "why is my Postgres connection pool exhausting"
Use GitHub issues, Reddit threads and support tickets to collect implementation questions. Treat their resemblance to private AI prompts as a research assumption, not an observed equivalence.
Run every engine you care about
ChatGPT with search, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude with web search each pick sources differently. Run each prompt on each engine.
Log what each answer cites
For every answer, record each cited URL, its domain, its position in the answer, and whether the answer named your product or a competitor. Then group domains into source types: official docs, GitHub, community forums, video, editorial, vendor blogs and review sites.
Repeat runs and report shares
Answers vary run to run. Fan-out searches differ by user and context, and one study found only about 27% of fan-out queries stayed the same across repeated runs.[9] Run every prompt several times and count the share of runs that cited each domain, not a single answer.
Keep claims honest
Treat the sample as descriptive evidence. A 2026 survey of 45 generative engine optimization (GEO) studies found no technique with a stable, cross-platform causal effect on organic discoverability.[10] Use the sample to identify relevant sources for further research. Presence on those sources does not guarantee a citation.
What should developer tool teams do with the map?
Prioritize documentation and community work using the source types observed in your category. Keep support needs and community rules ahead of citation targets.
- Identify frequently cited domains. Note which five domains carry most citations for your prompts.
- Review coverage on those domains. Search those domains for your product name. Count threads, issues and docs pages that mention you, and read what they say.
- Improve the source material. The measurement study found higher absorption for code, statistics, definitions and comparisons.[5] Make sure your docs and comparison pages carry all four.
- Respond to relevant open questions. Answer the open GitHub issue. Reply to the Reddit thread with a working example. Disclose that you work there.
- Measure monthly. Citation shares move fast. Check again next month.
For the mechanics of how engines pick sources, read Query fan-out, explained for developer questions. For the full playbook that connects listening and AI search, read Answer engine optimization and developer social listening.
Use these studies to form questions worth checking, then choose a fixed set of prompts for repeat measurement with the AI search prompt strategy guide. Its worked example connects each question with the evidence a useful answer would need.
Turn the source review into an assignment
If a competitor appears in a relevant answer, use the citation-gap audit to check product eligibility, source accuracy and page access. For an outside page that needs a correction or test, prepare the third-party contribution brief. For a missing owned reference or procedure, use the content-prioritization workflow to choose the smallest useful asset.
Sources
- AI Platform Citation Patterns. Profound.
- Most-cited domains in AI search. Peec AI.
- The Most-Cited Domains in AI: A 3-Month Study. Semrush.
- Reddit Is Winning the AI Game. Columbia Journalism Review.
- From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms. arXiv 2604.25707 (Zhang, He, Yao).
- An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied). Ahrefs.
- AIO Impact on Google CTR: 2026 Update. Seer Interactive.
- 2025 Stack Overflow Developer Survey. Stack Overflow.
- Ranking for Multiple Fan-Out Queries (173,902 URLs Studied). Surfer.
- Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026). arXiv 2607.14035 (Olivier Martinez).
Last verified October 22, 2026.