Most GEO (generative engine optimization) tools include a sources report: a ranked list of the websites that AI answers point to when they talk about your category. It is one of the reports that turns most directly into work. Teams use it to choose PR targets, plan outreach and find content gaps. Peec AI's Gap Analysis, for example, "highlights sources where your competitors appear, but your brand does not". Profound flags topics as "Visible, Not Cited" when "your brand is mentioned, but answer engines cite other sources", which points you to the sites that are cited instead.
The trouble is that a "top sources" list is not one standard measurement. Each tool builds it from four choices: which event counts as a source, how that event is counted, whether pages or whole sites are ranked, and how sites are labelled. Those choices can reorder the list. This article walks through each one, so you know what to check before you turn the list into an outreach plan.
All vendor details below come from each vendor's public documentation, read on 12 October 2026. They are documented capabilities or vendor claims. We have not independently tested any vendor's source report, and we know of no independent study that compares them.
Did the AI cite the site, or only look at it?
The first choice is what counts as a source at all. AI engines usually consult more pages than they cite in the answer.
OpenAI documents this for its developer API. Its web search guide separates inline citations, "which show only the most relevant references", from a sources field listing every URL the model consulted. It adds: "The number of sources is often greater than the number of citations." This describes the API, not the consumer ChatGPT app.
Peec AI, a GEO vendor, reports the same gap for Perplexity. In a February 2026 study of more than 1 million citations, it found that 64% of the Perplexity URLs in its dataset were never cited. As the study puts it, Perplexity shows those sources in its sidebar but is "much less likely to cite them in the actual AI answers".
GEO tools handle this gap differently:
- Peec AI counts two events separately. Sources are "the URLs AI models access while generating a response", while citations are "the sources explicitly referenced in the final answer". Its Domains report has separate Retrieved and Cited metrics. Its API also has a Links report for a third event: URLs the answer wrote into its text that are not among the pages it retrieved or cited.
- Genezio says that "citations" and "sources" mean the same thing in its product. The same page then separates "cited sources", which the engine explicitly cited, from "mentioned sources", which it "only referred to or showed, without an explicit citation".
- Profound, Semrush, Rankscale, AthenaHQ, Otterly.AI and Ahrefs describe only cited or linked URLs on the documentation pages we read.
Neither the study nor Peec's documentation explains whether its retrieval data comes from the consumer apps or from the engines' APIs. So do not assume retrieval data exists for every engine you track.
What this means: a list ranked by retrievals answers "what did the AI read?" A list ranked by citations answers "what did the AI point the user to?" Both are useful, but they are different lists.
Is the list ranked by answers or by citations?
Even when two tools count the same event, they can count it in different ways. Some count answers. Others count citations.
A per-answer metric asks: in what share of answers does this site appear at least once? Peec's "Cited", Profound's "Citation Coverage" and AthenaHQ's "Citation %" work this way. So does Semrush's "Citations", which its glossary defines as "the number of AI responses that cite your domain as a source". Genezio's "occurrences" are "the number of conversations that have the citation".
A per-citation metric asks: what share of all citations go to this site? A site cited five times in one answer gets five counts. Peec's and Profound's "Citation share" work this way. Peec's documentation spells out the consequence: Cited percentages across sources do not add up to 100%, but Citation share does.
For some tools, the pages we read don't settle the question. Rankscale lists the top URLs in each category with their citation counts, but doesn't say whether repeat citations within one answer count separately. Otterly's "domain coverage" and Ahrefs' citation counts don't say whether they count answers or citations either.
The two methods can put different sites at the top. Here is an illustration with invented numbers, not data from any tool:
| 100 tracked answers | Site A | Site B |
|---|---|---|
| Answers that cite the site | 30 | 10 |
| Citations per answer | 1 | 5 |
| Total citations | 30 | 50 |
| Per-answer rank | 1st (30%) | 2nd (10%) |
| Per-citation rank | 2nd | 1st |
Site A reaches more answers. Site B collects more citations. Which one tops your list depends on the tool's default metric.
Pages or whole sites?
The level of grouping matters too. A hostname is a full site address such as blog.example.com, while the root domain is example.com, covering every subdomain. Profound can report at either level. AthenaHQ's domain view "aggregates citations by root domain", and Genezio offers grouping by domain.
A large site with many lightly cited pages can lead a domain view and vanish from a page view. For outreach, that difference decides whether you pitch a publication or a specific article.
Who decides whether a site is "editorial" or "social"?
Once sources are counted, most tools label them by type. Some dashboards then show charts such as "share of citations from social or UGC (user-generated content)". Those labels come from each vendor's own scheme, and the schemes disagree.
- Profound puts review aggregators such as G2, Capterra and TrustRadius under "Social", next to Reddit and YouTube. It puts Wikipedia under "Institution". Domains with fewer than 100 total citations go into "Other".
- Peec AI classifies domains as Corporate, Editorial, Institutional, UGC, Reference, Competitor or Other. It also has separate URL types, where G2, Yelp and Crunchbase are examples of a "Profile".
- Otterly.AI uses 12 categories, including Encyclopedia, Video and Community/Forum. It says these "are set by the system and cannot be customized".
- Genezio classifies each citation on three axes: ownership, source category and website type.
- Rankscale mixes content types and platforms, such as blog, news media, YouTube and comparison sites.
- AthenaHQ uses only Owned, Competitor, Partner and Third-party, plus custom tags.
The Semrush and Ahrefs pages we read don't document a category scheme, so check in the product before relying on any category chart from them.
So G2 is "Social" in Profound and a "Profile" page in Peec. Most labels are assigned automatically. Peec, Profound and AthenaHQ let you override them, which means a category chart can also change after someone re-labels a few sites. We found no published accuracy figures for any of these vendors' automatic classification.
What this means: a category share is a vendor label applied to a count, not a measurement of the web. Compare category charts only within one tool, and only after checking how that tool labels your key sites.
Can you act on the gap flags and links as shown?
Classified lists usually come with flags such as "your brand is not mentioned on this source". These flags are where outreach plans start, so it helps to know what they are based on.
For Peec and AthenaHQ, the flag comes from the vendor's own later fetch of the page, which may differ from what the AI read. Peec lists which brands "were mentioned on this specific URL". Its "Updated" column is "the latest time we fetched content from this URL", which may be later than when the AI used the page. Peec says the text it fetched is "the same content that AI models read", but that is a vendor claim. AthenaHQ is the most explicit about limits. It hides its page-level flag in social views "because a page scan can miss mentions in comments". It also warns that rescanning a page "does not rewrite historical AI responses".
Otterly and Rankscale also show a brand-mention indicator for cited sources. Otterly marks each cited URL with "brand mention (Yes/No)", and Rankscale shows "which brands were actually mentioned within citations from that domain". Neither page documents how that is determined, or whether it reflects the page itself or the answers that cite it.
The cited links themselves can also be wrong. Ahrefs says it does "not filter out hallucinated or malformed links, as they reflect real model output". So a URL at the top of a list is worth opening before you plan around it.
Independent testing shows how often that can matter. A Tow Center study, published in March 2025, asked eight AI search tools which news article an excerpt came from, across 1,600 queries. Collectively, they answered more than 60% of those queries incorrectly. Some cited broken or made-up URLs, and others pointed to syndicated copies on sites such as Yahoo News instead of the original publisher. The study covered news articles in early 2025, not brand queries or today's engines.
What to ask before you plan outreach from the list
Each question below maps to one of the choices above:
- Event: Does the list count retrieved, cited or linked pages? Which one does the default view sort by?
- Counting: Is the ranking per answer or per citation?
- Level: Is it grouped by URL, hostname or root domain?
- Labels: Who assigns categories, can you override them, and where do small domains go?
- Gap flags: How is "brand mentioned on source" decided, and when was the page last scanned?
- Link quality: Are broken or hallucinated links filtered out or flagged?
- Collection: For each engine, does the data come from the consumer app or the API? Our comparison of GEO tools covers why that changes which sources appear.
- Evidence: Can you export the answers with their cited URLs, so you can spot-check the top rows?
The answers will not make one tool's list "right". They tell you which question the list answers. Compare source lists only within one tool, and read the answers behind the top rows before you commit outreach budget to them. For how source data fits with other metrics, see mentions, citations, sentiment and share of voice.