AI mentions, recommendations and citations: what each tells you

What each kind of evidence lets you conclude

Evidence	What it is	Where it appears	What you can conclude	What you cannot conclude
Mention	The brand name is in the answer text	Body of the answer	The platform associates the name with the question asked	That the brand was recommended, that the information is correct or where it came from
Recommendation	The answer puts the brand forward as the choice for the user's task	Body of the answer, read together with any conditions attached	The brand entered consideration with a defined role	That the recommendation will repeat, that the attributes given are accurate or that a source was involved
Citation	A URL is presented as a source for the answer or for a passage	References, numbered notes, links beside or below the answer	The URL was displayed as a reference, linked to a passage when the path reports that link	That the platform accessed the whole page, that it supports the passage, that it was used to generate the answer or that anyone clicked
Retrieved source	A page the platform reports having fetched during its search	Search data the platform exposes, when it exposes any	The platform reported the page among the results retrieved for that query	That it was read in full, that it weighed on the answer or that the user ever sees it

Mentions and recommendations are statements about the text: what the answer says about the brand and in which role. Citations and retrieved sources are statements about pages: what the platform reports having retrieved and what it displays as a reference. The two axes are independent, and a brand can sit in any combination: mentioned with none of its pages cited, cited without appearing in the text, recommended while a third-party page is displayed as a reference, retrieved without being cited. That is why the evidence does not form rungs ("first a mention, then a citation, then a recommendation") but a grid with two axes.

Mention: the brand is in the text

A mention is the easiest of the four to automate and the one most dashboards count when they talk about visibility. The count is legitimate as long as everyone knows what it leaves out. The name may have reached the text through the model's training data or through a page fetched while the answer was being built, and the text alone does not tell you which.

The brand's role is left out too. An answer can mention a brand as the recommended option, as one of eight alternatives, as a contrast ("unlike Brand X, which does not offer..."), as an example of what to avoid, or simply because the user typed the name into the prompt and the answer repeated it. To an automated count, all five situations are worth the same: one mention.

When the prompt already contains the brand name, a mention in the answer is close to guaranteed and says nothing about unprompted presence. Category prompts ("which help desk software suits a ten-person support team?") measure whether the brand appears on its own; branded prompts ("what is Brand X?") measure something else, namely how the brand is described. Mixing the two kinds in one percentage inflates the mention rate without anything having changed on the platform.

Recommendation: the answer picks the brand

Recognizing a recommendation takes three criteria, and none of them is the presence of the name. The answer has to be addressing the user's task rather than a side topic. The brand has to be presented as a choice, on its own or within a short set, rather than as an item in an inventory. And the conditions attached to the pick ("if you need telephony integration, Brand X is the better fit") have to be read together with it, because a recommendation tied to a scenario that is not your buyer's has a different commercial value. A passage such as "for small teams, Brand X; for multilingual operations, Brand Y" recommends both, each for one scenario, and the report needs to preserve that.

None of these criteria can be checked by counting names. Recommendation requires reading the answer, and a tool that reports a "recommendation rate" is applying some classification rule to the text, human or automated, which needs to be explained before that number is compared with any other. Lastfind's mention metric is detection of the brand name in the final answer text and does not classify sentiment or recommendation; deciding whether the brand was recommended is a reading of the archived answer.

Two weaknesses come with any recommendation and belong in the record. Without a citation, you cannot distinguish whether the answer came from training, an undisclosed search or another path, and nothing guarantees the next run will repeat the recommendation. A recommendation with wrong attributes is fragile too: correcting public sources lets you test whether the description changes, but does not guarantee that it will.

Citation: a page appears as a source

The form of a citation varies across platforms: a bracketed number tied to a list, a card with title and domain, a link embedded in the sentence, a reference list at the end. All of them refer to a page, not to a brand. The cited page may belong to the brand's site, to a competitor, to a publication, to a forum or to a marketplace.

In programming interfaces the link is explicit. In OpenAI's Responses API, each citation is a url_citation annotation carrying the URL, the title and the location of the cited passage, and OpenAI requires it to be visible and clickable when the answer is shown to end users. Anthropic's API with web search returns, for each citation, the URL, the title and up to 150 characters of the cited text, and citations are always enabled when search is used. In the Gemini API with Google Search grounding, each url_citation annotation links a text segment, delimited by start and end indexes, to a source URL. In all three, a citation is a pair: a passage of the answer and the page the platform associates with it.

That is different from a mention: it shows that the platform displayed a URL as a reference and, when the contract includes indexes, linked it to a passage. Anyone can open the URL and check its content, but the link alone does not prove access to the whole page, use in generation or support for the passage.

What is usually read into a citation goes beyond what it proves. The audit of four generative search engines published by Liu, Zhang and Liang in 2023 found that, on average, 51.5% of generated sentences were fully supported by their citations and 74.5% of citations supported the sentence they were attached to. The study covers Bing Chat, NeevaAI, Perplexity and YouChat; the systems have changed since, but the point stands: a citation is a link the platform presents, not a verification that the page's content backs the passage.

The click is not in the citation either. For Google, sites appearing in AI Overviews and AI Mode are included in the overall search traffic in Search Console and reported in the Performance report within the "Web" search type; seeing your page cited does not tell you how much traffic the citation produced.

Retrieved source: the platform fetched the page, which does not say it used it

It is the newest of the four kinds of evidence and the most misread, because it looks like it says "the AI read this". In retrieval-augmented generation as described by Lewis and colleagues in 2020, the system retrieves passages from an index and generates text conditioned on them; the architecture does not fix how many passages are retrieved or which ones reach the final text. In implementations that expose sources, retrieval can return more URLs than the citations displayed.

At OpenAI, the sources field returns the complete list of URLs the model consulted when forming its response, while inline citations show only the most relevant references; the number of sources is often greater than the number of citations. Anthropic returns the results of each search in a block of their own, separate from citations. In the Perplexity Agent API web search tool, the response can include a search_results output item listing the results, each with ID, URL, title, snippet and dates.

For anyone who works on content, that list separates two hypotheses that a citation alone blurs: the page may not have been retrieved, or it may have been retrieved without appearing as a citation. The list does not show why a page was not cited or whether it influenced the final text. The evidence holds for that query, in that index, on that day; it is not a permanent property of the page.

A retrieved page may have influenced the answer without receiving a citation, may not have influenced the final text, or may be cited and still contradicted by the text. The phrase "the AI read the site" is not supported by any retrieval list.

There is also a third state that tends to stay invisible: the platform may not have searched at all. In Anthropic's API, the model decides from the prompt whether to search and answers directly, without a search, when the question rests on stable knowledge. An answer with no retrieval list can therefore mean that the platform searched and did not expose the list, that it searched and the list came back empty, or that it did not search. A serious record keeps that difference: "empty list" and "not reported" are distinct observations, and only the first lets you say the search found nothing.

Six cases that mislead

The cases below are hypothetical. Imagine your company sells help desk software and the monitored prompt is "which help desk software suits a ten-person support team?".

The brand's page is cited as a source, but the brand is not recommended

The answer recommends a competitor and displays, as a reference beside its price comparison, a page on your own blog comparing the two tools. Your brand is in the text, your page is in the references, and the automated report logs a mention and a citation. That is where the observation ends: the citation does not prove that the platform accessed the whole page or used its content to choose the competitor. The editorial decision is to open the URL and check whether the page supports the comparison and describes the competitor with more precision, attributes or recent information than it gives your own offer. Finding that difference is a hypothesis for the test, not a conclusion from the citation.

The brand is recommended without its site being cited

The answer names your brand as the best option for small teams and displays a review on a ratings site and a forum thread as references. No page from your domain appears. The recommendation exists; the references identify pages worth opening, but do not prove that they caused or support the choice. The decision is to check whether those pages describe your brand accurately, request legitimate corrections where needed and publish on your own site information that today exists only elsewhere. New collections show whether the answer or references change, without attributing that change to one page.

A competitor is cited in the answer that recommends you

A variant of the previous case: the answer recommends your brand for ten-person teams and displays as a reference the competitor's comparison, which positions its product for larger operations and concedes the small segment to you. The recommendation and reference are observable; the citation does not prove that the page determined the framing. The decision is to publish your own comparison with verifiable criteria and watch in new collections whether it is retrieved or cited, without assuming replacement or causal effect.

The brand appears only in a list of alternatives

The answer recommends two tools, explains why, and closes with "other options include" and six names, yours among them. That is a mention, not a recommendation. It shows that the platform associates your brand with the category, but no attribute that would justify choosing it. The decision this case supports is about positioning: find out which attributes the two recommended tools were given in the text (price, integrations, language, support) and check whether your content and third-party sources attach any choice criterion to your brand, or only the name and the category.

The brand appears with wrong information

The answer mentions your brand and states that it "has no free plan", when it does. If a citation is attached to the passage, open the page and check whether it repeats the outdated information; it may be a review, a third-party comparison or one of your pages. If there is a retrieved-source list, check whether an outdated page from your domain is on it, without inferring that it was used. With neither a citation nor a reported retrieval, the cause remains unknown: training, an undisclosed search or another path may be involved. In any scenario, correct public sources and collect again to watch whether the description changes; none of them guarantees an immediate effect.

The brand's page is retrieved but neither cited nor mentioned

The retrieved-source list includes your pricing page, and the answer recommends two other tools without mentioning you. The evidence shows that the page was retrieved in that run but was neither displayed as a reference nor associated with the brand in the final text; it does not show why. This case appears on no dashboard that counts only mentions and citations, and it offers a useful hypothesis for anyone producing content: compare your page with the ones that were cited, looking at which questions they answer, which attributes they name and which format they use. Those differences can guide the investigation, but the collection alone does not establish that content was the cause.

There is also the case in which none of the four kinds of evidence appears: the brand is not mentioned, and nothing from its domain is cited or retrieved. The question then stops being "how does the AI treat us" and becomes whether any public content answers the prompt with the brand name; the decision starts there.

In every case, the decisions are about content, relationships with sources and positioning, which is the work of GEO as a front within SEO; none of them is "ask the AI to change its answer". The collection does not decide by itself whether acting is worth it either: a recommendation in prompts that do not match the buyer's journey, or on a platform the company's audience does not use, can be logged and set aside. Priority comes from the business; the evidence only makes the decision possible.

Four counts from the same collection

The example is hypothetical. Suppose 40 valid answers to the same set of category prompts. Your brand is mentioned in 30 of them, recommended in 8, has a page from its own domain cited in 5 and reported as retrieved in 22. For this calculation, the 5 answers with a citation are a subset of the 22 with a reported retrieval.

Evidence	Answers	Rate (÷ 40)
Mention	30	75%
Recommendation	8	20%
Own page cited	5	12.5%
Own page retrieved	22	55%

The four rates describe the same collection and tell different stories. A presentation that says "75% visibility" is correct and says little about purchase decisions; the recommendation in 20% of answers is the number that matters to whoever sells.

Because the example declared the subset relationship, 22 − 5 = 17 answers in which an own page was reported as retrieved but not displayed as a reference; that defines a queue of pages to inspect, not proof of use. The 30 mentions include the 8 recommendations and, very likely, a few lists of alternatives and the odd contrast; only reading separates those roles. None of the four counts is share of voice, which compares brands with one another and has a different denominator.

What changes by platform

Platforms expose different evidence, and the same platform exposes different evidence depending on the collection path. The table describes the documented contract of each API. A chat interface can expose different evidence and changes often, and what a monitoring tool captures depends on its collection path.

Platform	Citations	Retrieved sources	Note
OpenAI Responses API with web search	url_citation annotations with URL, title and passage location	sources field, with the complete list of URLs consulted, distinct from citations	API contract; the ChatGPT interface may expose different evidence
Anthropic API with web search	Always enabled when search is used; include up to 150 characters of cited text	Results of each search returned in their own block; a search with no matches returns an empty list	API contract; the Claude interface may differ
Google Gemini (API)	url_citation annotations linking passages to URLs	Queries the model executed; no documented list of fetched pages beyond the cited ones	An executed query is a third kind of evidence: it says what was searched, not what was found
Perplexity (Agent API)	No citations field, declared "fully deprecated and removed" in a May 2025 breaking change; inline citation markers in the answer text, tied to the id of each result, depend on the prompt or the preset used	search_results output item, with the search queries and a results list (ID, URL, title, snippet, date, last updated, source)	Sonar Chat Completions is now the Agent API: Sonar support ended on September 27, 2026, and requests are reformulated as Agent API requests; whether a result was actually used in the answer is not documented
Google AI Overviews and AI Mode	Links shown with the answer; the set of links can differ between the two features	Not exposed; Google describes "query fan-out", several searches across subtopics, without publishing the list of pages	AI Overviews often do not appear; the organic results beside them are not citations of the answer

Two practical consequences. First, a "retrieved source" exists only where the path reports it: where search may occur, an absent list does not distinguish an answer without search from a search whose results were not exposed, and retrieval before generation as Lewis and colleagues describe it holds for the RAG models they studied, not for every platform. Second, collection defines the evidence: an API returns its documented fields and an interface offers what the screen shows. Comparing citation rates across platforms or paths is confounded by both exposure mechanisms and system behavior; this design isolates neither one.

Contracts also change. Perplexity removed the citations field in favor of search_results and, in July 2026, announced that Sonar Chat Completions is now the Agent API. A historical series of citations needs to note the date on which the meaning of the field changed.

What a record needs to keep to tell the four cases apart

Any monitoring process, manual or automated, can separate the four kinds of evidence if it keeps five things:

• The complete original answer, because mention and recommendation can only be told apart by reading the text, and a citation can only be checked by opening the page the answer pointed to.
• The citations as a list of their own, with the URL and, where available, the passage each one was attached to.
• The retrieved sources as a second list, separate from the first, with the distinction between "the platform returned an empty list" and "the platform did not report".
• The collection path and the date, because the meaning of the fields changes.
• The exact prompt, because a prompt that already contains the brand name makes the mention trivial.

Whoever keeps only the final percentage cannot go back to any of the six cases above. A citation logged without the answer does not tell you whether the brand was recommended; a mention rate without the prompt does not tell you whether the presence was unprompted; a source list without the empty-versus-absent distinction does not tell you whether the search failed or was never run.

Lastfind keeps cited sources and retrieved sources separate, according to what each provider reports, and preserves the distinction between an empty list and unreported data. AI source and citation analysis starts from that separation. Each original answer is archived with an integrity proof and can be opened from the record: recommendation and context are read in the real text, not computed by an automatic metric, and that is what makes the numbers checkable.

A test for your next collection

Take ten answers from your latest collection, preferably from category prompts in which the brand appeared. For each one, fill in four columns with yes or no: mentioned, recommended, own page cited, own page retrieved (or "not reported", where the platform exposes nothing). Then compare the columns with what the dashboard reported.

If the ten rows under "mentioned" match the ten under "recommended", the coincidence may be real in that sample; inspect each answer and the dashboard rule before concluding that it classifies recommendations by an unknown criterion or that you are reading mentions as recommendations. If "cited" and "retrieved" are always identical, that equality may also be sample-specific; inspect the fields before concluding that the record does not separate the lists. And if you cannot fill in a column because the original answer was not kept, that is the first problem to fix, before any discussion about the rate.

Platform documentation checked in October 2026.