AI citations are not recommendations

A company can win an AI citation and still lose the buyer. Imagine an answer to “What should a 20-person agency use for client reporting?” that links to your guide explaining reporting workflows, then recommends a competitor because it says that competitor handles client approvals better. Your URL appeared. Your brand may even be named. The decision still went somewhere else.

The opposite is common too. An assistant recommends a familiar brand without visibly linking to its site. That may be commercially valuable, but it gives the marketing team no owned page to improve and no evidence that a new article earned the mention.

AI citations, brand mentions and recommendations are three different events. They sometimes arrive together. They do not have to. Combining them into one score makes it impossible to tell whether you need a better source, a clearer reason to choose, wider brand evidence or a product decision.

For practical reporting, count an AI citation when the answer visibly attaches a source link to a page or domain. Record the actual URL, the sentence it appears to support and the assistant surface where you observed it. A domain listed in a general sources tray is still useful to record, but do not assume it supports every statement in the answer.

ChatGPT's help center says responses that use web search may include citations and that the citations can be incomplete, outdated or incorrect. Its guidance tells readers to open the source and check it. Treat your tracking record the same way. A citation is an observable pointer, not a certificate that the cited page caused the surrounding wording.

The OpenAI API distinguishes inline citations from a larger sources field in web-search responses. That distinction is useful for people building their own measurement, but it does not mean every consumer chat exposes the same evidence. Record what the interface actually showed.

A recommendation answers the buyer's selection question. The answer says, in substance, that this product fits the stated need, ranks it above alternatives, or includes it in a shortlist with a reason that applies to the prompt. It is not enough for the assistant to mention your name in a history of the category.

Score the recommendation against the condition in the prompt. If the buyer asks for an affordable tool with a required integration, “Beseen is good for larger teams” is not a recommendation for that request. It may be an accurate description. It still fails the selection test.

A brand mention is simpler: your company name appears in the generated text. That can be a useful early signal, especially for discovery prompts where a buyer is collecting options. It becomes less useful when the wording is negative, inaccurate or detached from the buyer's need.

Keep the answer excerpt with the label. “Beseen is often used for agency reporting” and “Beseen may be too limited for this workflow” should not become the same green cell in a spreadsheet. The former may be a neutral mention. The latter may expose an explanation, positioning or product gap worth investigating.

This also explains why ordinary web mention tools cannot substitute for answer sampling. A public review, forum post or article can help shape what is available to an assistant. It is not evidence that the assistant named your brand in a private, generated response.

Your page is cited, but the competitor is recommended

Inspect what your page contributed. It may define the category while the competitor's page provides the decisive capability. A better next step may be a comparison with sourced trade-offs, not more category copy. Comparison pages need a reason to choose, not a column of unchecked superlatives.

Your brand is recommended with no owned citation

This can reflect a well-known brand, third-party coverage or a response whose evidence is not visible. Do not claim your latest page drove it. Record the recommendation, inspect any visible sources and look for the buyer condition that made the assistant choose you.

Your domain is cited, but your brand is not named

The page may be supporting a factual explanation rather than a vendor choice. That can still bring useful visits. It is not proof that your company entered the shortlist.

Your brand is named inaccurately

A mention is not a win when it repeats the wrong price, scope or identity. Save the statement and check its evidence. The repair may belong in an owned source, a partner listing or a clarification of product boundaries. The correction workflow begins with the claim, not with a new homepage rewrite.

Run a fixed set of buyer prompts under comparable conditions. For each valid answer, store the full text and score the outcome separately. Use a row per run, not a row per month. A summary can then calculate the rate without hiding the answers that produced it.

• Brand named in the answer body: yes, no or unclear.
• Brand recommended for the stated constraint: yes, no or unclear.
• Owned URL visibly cited: yes or no, with the exact URL.
• Competitor named or recommended: record the name and reason given.
• Accuracy of each material claim: correct, incorrect, unverified or not stated.

Keep “unclear” as a real outcome. An answer can list products without selecting one. A source list can be visible without a clear attachment to a sentence. Forcing a binary label creates precision that was not present in the answer.

You generally cannot establish that from a single answer. A citation can be adjacent to a recommendation, support one fact inside it or be one of several sources the assistant considered. The reply is generated, and the visible citation is not a complete account of its internal process.

That does not make citations useless. They offer a concrete lead: open the source, see what it says, compare it with the claim and check whether the page is accessible and current. Google's review guidance also favors original evidence, meaningful differences and explaining what circumstances favor a choice. Those are sensible qualities for a page a buyer needs, regardless of what a single assistant did with it.

If you change a page after a citation gap, take a baseline and repeat the same prompts. Do not confuse a later citation with proof of causality. A before-and-after study needs repeated runs and a comparison before it can support a stronger claim.

Consider a fictional scheduling platform called Harbour. A buyer asks which scheduling tool suits a small clinic that needs two-way calendar sync and does not want a ten-seat minimum. In ten valid runs, Harbour's pricing page is cited four times. Harbour is named six times. It is recommended twice. A competitor called Maple is recommended seven times because the answer says its entry plan covers the team size.

Calling Harbour's four citations a win would hide the buyer's actual decision. Calling six mentions a 60% recommendation rate would be worse. The useful report says that Harbour appears, its pricing source is sometimes available to the answer, and the seat requirement is still directing most selections to Maple. That points the investigation at a product fact and the explanation around it.

Now imagine that two of the six Harbour mentions say it has a ten-seat minimum when the current plan starts at two. That does not improve the success count. It creates an accuracy issue. The record should show two incorrect claims, two recommendations, four citations and the saved answers behind each result. Every outcome gives a different owner something concrete to check.

Open the cited page and find the text closest to the claim. A citation next to a recommendation may point to a review, a documentation page or a third-party list. If it describes the competitor accurately, the result may simply reflect a real difference. If it contains an old limit or an unsupported comparison, save the passage before asking anyone to change it.

Keep three statements separate in the investigation: what the assistant said, what the cited source says and what you think led to the answer. The first two are observable. The third is a hypothesis. This sounds fussy until a team starts editing pages based on a source that never made the disputed claim.

Check your own cited pages the same way. A citation to an old launch post is not a reason to rewrite every historical page. It may need a current-terms link or a tighter statement about the date and plan it described. Correct the smallest demonstrated contradiction, then retest the prompt later.

The first mistake is counting every source from an answer as a source for your brand. If the assistant shows six links, only one may concern your company. The second is treating a citation to your home page as proof that the homepage answered the question. A page can be visible without doing useful work in the decision.

The third is measuring only the answer that has a citation. Search and answer modes can differ, and a buyer may ask a follow-up that removes the source links. Label the mode and save the full thread context that matters. A change in mode is a change in the measurement, not a normal fluctuation.

Finally, do not optimize for citation count by creating pages that repeat the same shallow claim. A buyer still needs a clear condition, evidence and a reason to choose. A page that is briefly cited for a generic definition may add less commercial value than a carefully maintained comparison that helps the right buyer rule your product out or in.

Missing citations and missing recommendations lead to different work. When an owned page is missing but the product is already recommended, check whether a helpful current source exists for the claim, whether it is reachable and whether it says the condition close to the fact. When the page is cited but the competitor wins, inspect the selection reason. The constraint may expose a real product gap or a comparison that does not answer the buyer question.

When the brand is missing entirely, resist the instinct to publish generic “AI search” content. Find the prompts where the buyer problem overlaps your product, the rivals named and the evidence behind the answer. Sometimes the best response is a clear product page. Sometimes it is a third-party factual correction. Sometimes the answer describes a capability you do not offer.

A clean summary might say: “Beseen was cited in seven of twenty valid evaluation runs, recommended in four and named in nine. In five of the seven cited runs, Quarry was still the recommendation because the prompt required exportable dashboard layouts. Two Beseen descriptions contained an unverified pricing claim.” This tells a team where to look without pretending that citation rate is revenue or recommendation rate is a ranking.

Link the numbers to the saved answers. A stakeholder should be able to click from the report to the exact prompt, result and source without asking an analyst to recreate it. The article on building an AI visibility report covers the wider structure, including prompt versions and competitor context.

AI citations are valuable observations. Recommendations are valuable observations. Brand mentions are valuable observations. The mistake is not tracking any of them. It is claiming that one convenient count answers every question about discovery, trust and purchase intent.

Beseen records the prompts, names, competitors and visible evidence behind the gap, so your team can separate a missing source from a lost recommendation. Start with a free report, then make the smallest change the evidence supports. The point is not to win a dashboard. It is to help the next buyer reach an answer you can stand behind.