A Citation Is Not a Recommendation

Two studies, one metric, a four-fold gap

In 2026 two teams published research on the same narrow question: when an AI engine cites your page, how often does the answer also say your brand's name out loud?

Semrush, working with Kevin Indig, logged 3,981 domain appearances across 115 prompts, 14 countries and four engines. BuzzStream logged 12,000 AI responses covering roughly 200 brands across 10 industries, resolving 221,946 cited URLs of which 47,550 belonged to tracked brands.

Semrush published three buckets that sum to every appearance in its dataset: 61.7% cited with no brand mention, 13.2% cited and mentioned, 25.1% mentioned with no citation. Citations therefore account for 74.9% of appearances. Divide the 13.2 by the 74.9 and the share of citations that also name the brand is 17.6%.

BuzzStream published that exact ratio directly. It got 69.9%.

Same metric. Four times apart. No arithmetic error on either side.

The temptation is to call this a methodology dispute and move on. It is not. Both teams measured carefully and documented their samples. What they did not share was a definition, and the field has never forced them to. That is the actual problem, and it is a vocabulary problem, not a data problem.

Nobody agrees what a citation is, and Google does not use the word

I fetched the pages currently ranking for this topic and pulled their definitions verbatim. They do not agree, and the disagreement is not cosmetic.

Read the third column again. Similarweb defines an AI mention as your brand appearing "without linking to your content," which means a linked brand name is not a mention. BuzzStream counts mentions "with or without a link," which means the identical event is a mention. Two pages ranking for the same query give opposite answers about the same pixel on the same screen.

ZipTie is stranger still. It classifies by evaluative framing rather than by linking, so a bare source link in the tray is not a citation at all under its rule, while a sentence in the answer body can be. That is a defensible taxonomy. It is simply not the one anyone else is using, and nobody says so.

This is what a young field looks like before it standardizes. Ahrefs' Louise Linehan describes the last twelve months as a shift from counting mentions to interrogating their value, one that has spawned an entire measurement vocabulary. She is right about the direction. The problem is that the vocabulary arrived faster than the definitions did, and vendors shipped dashboards on top of it anyway.

The cleanest evidence that the industry invented this vocabulary is that the largest engine does not share it. Google's AI features documentation never uses the word citation. It says pages are eligible "to be shown as a supporting link in AI Overviews or AI Mode." The noun the entire measurement category is built on does not appear in the primary vendor's own spec.

The four events, with a counting rule for each

Here is the taxonomy I use and the one the rest of the measurement posts on this site run on. Four events, four counting rules, four different commercial meanings.

1. Citation. Your URL appears in the answer's source apparatus: tray, sidebar, footnote or inline link. Count one per answer per domain, no matter how many links point at you. It does not require your name to appear anywhere a human reads. Commercial value is the lowest of the four, because Pew's passive panel of 900 US adults found users clicked a link inside an AI summary in just 1% of visits to pages that had one.

2. Mention. Your brand name appears in the answer text. Count one per answer, not one per repetition. This matters more than it sounds, and I will come back to it.

3. Inclusion. Your brand appears inside an enumerated set of options offered for the same job. Record two values, never one: the size of the set and your ordinal position in it. "In the set at four of six" is a materially different fact from "in the set."

4. Recommendation. The answer names your brand as the choice, with endorsing language directed at the user's stated need. Count it only when the endorsement is actually directed at that need, not when your brand is the example in a definition.

Use this graphic on your site

Free to republish with a link back to this page. Copy the embed code:

<a href="https://josephtimpson.com/insights/ai-citations-vs-recommendations"><img src="https://josephtimpson.com/assets/infographics/ai-citations-vs-recommendations.svg" alt="Four layer stack showing citation, mention, inclusion and recommendation as separate countable AI answer events, ordered by commercial value" width="1200" style="max-width:100%;height:auto"></a><p>Graphic by <a href="https://josephtimpson.com/insights/ai-citations-vs-recommendations">Joseph Timpson</a></p>

The ordering is by commercial value, not by nesting. You can be recommended without being cited, which happens constantly on Gemini, and cited without being named, which is the majority case on ChatGPT. Semrush's per-engine split makes that concrete: ChatGPT ran an 87% citation rate against a 20.7% mention rate, while Gemini ran 21.4% citations against 83.7% mentions. Those are dataset-wide rates rather than per-brand ones, so do not read them as your own odds. The direction is still unmistakable. The two engines make opposite default choices about attribution and naming, which means a dashboard that averages them is reporting on a system that does not exist.

That sentence is the reason event one and event two have to be separated. The engine treats attribution and naming as unrelated decisions. Any metric that fuses them is fusing two independent choices the model made for different reasons.

French, quoted in BuzzStream's study, gives the most useful mechanistic framing published so far, and it upgrades the taxonomy from bookkeeping into diagnosis. An unlinked mention is the model recalling you. A citation with no mention is the model checking something and not caring who you are. Those two failures need completely different fixes, and a blended visibility score hides which one you have.

Your prompt set decides your number, not the engine

This is the likeliest explanation for the 17.6 against 69.9 gap, and it is uncomfortable for anyone selling a dashboard.

BuzzStream segmented overlap by prompt type, running the ratio the other way round: the share of mentions that arrived with a citation. For single-brand prompts, where the user already named the company, 39% of mentions came with a citation. For open list and category prompts, the kind a real buyer asks, only 7.2% did. Head-to-head comparisons landed between them at 35.8%. Same engines, same brands, same window. That overlap moved more than five-fold on prompt selection alone.

Neither team published its prompt mix in a form that lets you recompute the other's number, so I cannot close the four-times gap arithmetically and will not pretend to. But a variable that swings one overlap measure five-fold is the first place to look, well before you blame the engines. Two agencies reporting a "citation to mention rate" to two clients are not reporting the same quantity, and neither client can tell.

SparkToro's volatility work compounds it. Across 2,961 prompt runs from 600 volunteers, the engines almost never returned the same brand list twice.

The supply of citations is not stable over time either, which quietly breaks month-over-month reporting. Seer Interactive's John Lovett analyzed 206,412 ChatGPT responses across 43 brands and found citations per response jumped from 5.7 to 10.4, an 81% increase, starting on a single date in December 2025. Any brand tracking raw citation counts booked a large win that month for doing nothing. That is why a defensible prompt set and an honest sample size matter more than the engine you pick, and why tool comparisons are meaningless until you know which of the four events each tool counts.

The loss that gets reported as a win

Now the commercial argument, which is the whole reason the taxonomy is worth the trouble.

Picture the answer to "best commercial roofer in Salt Lake City." Your page is in the source tray. The answer body names a competitor as the pick and lists two others as alternatives. Your brand name never appears in the text a human reads.

Every tool on the market records that as a citation. Most blend it into a visibility score. The score goes up. You lost the sale.

The cited-and-unnamed half of that is not an edge case. It is the largest bucket Semrush measured: 61.7% of appearances were a source link where the brand name never appeared in the answer. Note the denominator carefully, because the number is widely misquoted. 61.7% is the share of all domain appearances, not the share of citations. Run it against citations only and 82.4% of cited appearances named nobody. The correction makes the finding worse, not better.

Position inside the answer is its own separate event, and Peec AI is the only team I found measuring it at scale: roughly 200,000 responses across eight engines, September 2025 to March 2026. Ranking first inside a frequently cited third-party listicle moved a brand's position in the answer earlier by 1.80 places in US Finance, 1.17 in B2B SaaS and 0.82 in Emerging MarTech.

That is the strongest argument against frequency-based scoring I have seen, and it is why my counting rule for a mention is one per answer. Any tool that counts repetitions rewards being the cautionary example. Ehrlinspiel's own instruction is to track visibility and position as two separate metrics, which is the same conclusion arrived at from the other direction.

There is also a supply-side reason recommendations show up where you did not ask for them. Profound ran 50,000 prompts across seven industries and found that nearly half of every AI response contained unsolicited editorial content: comparisons, opinions and recommendations the user never requested. Roughly 47% of the content in verbose responses was material nobody asked for. Your brand is being ranked in answers to questions that were not ranking questions.

What to report instead

Replace the single visibility score with four lines. This is the format the rest of the measurement work on this site uses.

Report all four per engine and never averaged across engines. Averaging ChatGPT's 87% citation rate with Gemini's 21.4% produces a number describing no system that exists. Semrush's index work separately found the overlap between mentioned brands and cited domains on Gemini can run as low as 30%, with ChatGPT averaging about 15 sources per response against Gemini's three. These are different instruments pointed at different things.

The practical order of operations does not change much. Confirm crawler access first, because a blocked fetcher zeroes all four events at once. Understand that retrieval happens at the subquery level, which is why citations arrive from pages you were not targeting. Then work on the events themselves: earning citations is a different job from earning brand mentions, and both are different from being described correctly when you do get named, which is its own accuracy problem.

What changes is what you promise. A retainer that promises "AI visibility" has promised nothing measurable. A retainer that promises to move mention rate on a named prompt set, sampled at a stated cadence, has promised something a client can check and something you can fail. That is a better business to be in, and it is the only version of this work I am willing to sell.

These four terms are load bearing across everything else I publish. The MEASURE stage of the Cited Method is built on them, and every post in the insights archive uses them in exactly this sense. If you see me write "citation" anywhere on this site, I mean event one, and nothing more than event one.