Citation Growth Does Not Mean Sales Gets Named

A GEO vendor's monthly report says citations are up. Before that number goes into a sales deck, ask one question: how many of those citations put the brand's own name in the sentence a buyer reads? Most of the time, the honest answer is fewer than one in ten.

That is not a guess. The Machine Relations Index just measured it directly, and the measurement is a worksheet, not a caveat.

The number a citation count is hiding

Across 14,447 citations the Index could read in full, on four AI engines over a 139-day window, the cited source's own name appeared in the answer's prose 8.6% of the time. The other 91.4% of citations were silent: the engine used the page, attached it in a structured source layer, and never said whose page it was in the sentence the buyer actually read.

That gap is not evenly spread. It splits by engine and it splits by domain, and a sales team briefed on "citations" without either breakdown is being handed a number that cannot predict whether a prospect ever hears the brand's name.

By engine, on the same prompts in the same window:

Engine	Citations read	Named in the prose	Rate
Gemini	2,865	377	13.2%
ChatGPT	2,261	238	10.5%
Claude	1,820	184	10.1%
Perplexity	7,501	440	5.9%
Four engines pooled	14,447	1,239	8.6%

A brand cited only on Perplexity and a brand cited only on Gemini can post the same citation count and get named 2.2x as often on one of the two. A vendor report that sums citations across engines is adding two different outcomes together and calling the total one metric.

By domain, the spread is far wider than the engine split. Machine Relations tested 17 domains cited 100 times or more and compared each one's naming rate against its own not-cited control (how often the same brand name turns up in answers that never cited that domain, which is the baseline noise floor):

Domain type	Example	Named in the prose	Control rate
Product a buyer can choose	databricks.com	28.6%	0.6%
Product a buyer can choose	hubspot.com	21.3%	1.5%
Reported-on entity	gartner.com	11.9%	1.2%
General-purpose evidence	reddit.com	3.5%	1.0%
Third-party press evidence	forbes.com	2.1%	0.1%
Third-party press evidence	nytimes.com	0.0%	0.0%
Wire distribution	prnewswire.com	0.0%	0.0%
Academic evidence	arxiv.org	0.0%	0.0%

The pattern behind the numbers: a domain gets named when the engine is recommending it as the thing the buyer might choose. A domain gets cited but not named when the engine is only using it as evidence for a claim about someone else. Press coverage, wire distribution, and reference pages fall almost entirely into the second bucket, no matter how many times they are cited. A brand's own site and product pages, when the engine cites them at all, sit in the first.

That is the fact a PR or GEO vendor report rarely states, because their instrument counts a citation event, not the sentence built on top of it.

The worksheet: what to ask a GEO vendor before the number reaches sales

I have watched a marketing team walk a citation chart into a pipeline review and watch sales ask, correctly, "so did anyone actually hear our name?" The chart cannot answer that question as built. Five questions can, and every one of them is answerable from data a vendor already holds if they are asked for it.

1. Break the count by engine, not pooled. A pooled total hides a 2.2x spread. If the vendor cannot produce a per-engine table, the pooled number is not usable for a forecast.
2. Ask what share of the citations are to owned domains versus third-party coverage of the brand. Owned-domain citations run far closer to the 20-to-30% band; third-party evidence citations run closer to zero, whatever the outlet's prestige.
3. Ask for the not-cited control rate for the brand's own name on each engine. Without it, there is no way to tell a real citation effect from ordinary brand-name frequency in that engine's writing style.
4. Do not accept "citations" and "mentions" as the same metric inside one report. If a vendor's dashboard uses both words for one number, that is the tell that the distinction has not been built into the instrument.
5. Set the sales-facing target on naming rate, not citation count, and track it separately going forward. A citation count that rises while the naming rate stays flat is not evidence the brand is more visible to a buyer reading the answer; it may only mean the brand is cited by more evidence-role pages that were never going to say its name.

What this changes about the pipeline conversation

None of this argues against earning citations. A citation is a precondition: a domain has to be cited before it can be named, and the Index's own control shows a citation raises the odds of the name reaching the prose by a factor of 9 to 39, depending on the engine. The argument is narrower and more useful than "citations don't matter." It is that a citation count, reported without an engine breakdown and a domain-role breakdown, cannot tell a sales team what a prospect actually read.

The fix costs nothing to implement and nothing to measure: ask the GEO vendor for the same two cuts the Index publishes, by engine and by domain role, before the monthly number goes into a deck. If they cannot produce either cut, that is itself the finding to bring back to the vendor conversation, not to sales.

FAQ

Does a rising citation count mean anything? It means the brand's pages are entering more AI answers as sources. It does not, on its own, mean a buyer read the brand's name. The two events are correlated but the gap between them runs from 91% down to as low as 71% depending on engine and domain role, per the Machine Relations Index measurement above.

Which engines name a cited source most often? In this measurement, Gemini named a cited source in 13.2% of citations, the highest of the four engines with enough answer-text coverage to test; Perplexity named one in 5.9%, the lowest. Google AI Overviews and AI Mode were not part of this specific answer-text panel.

Why does a brand's own domain get named more than a press mention of the brand? Because engines name a source when they are recommending it as the answer to the buyer's question, and name it far less when the source is only evidence for a claim about a different subject. A product page the engine is recommending gets named; a news article the engine is citing as background evidence usually does not, even at high citation volume.

What should replace "citation count" as the number sales sees? Naming rate by engine, alongside the count, and a note on whether the citations driving the count are to owned domains or to third-party evidence about the brand. Track both quarter over quarter rather than the count alone.

Where does this data come from? The Machine Relations Index, release of September 26, 2026, drawing on a full per-engine breakdown published the same day. A companion piece on the buyer-vetting side of this same finding, aimed at what to demand from a GEO vendor before trusting their citation-count metric, is at authoritytech.io. Source attribution mechanics for the four engines: OpenAI's web search tool, Anthropic's citation blocks, Perplexity's citation quickstart, and Google's documentation on AI Overviews and AI features in Search.

---

This is the same source-role distinction Machine Relations tracks in its own Index measurement, and the same discipline AuthorityTech applies when it reports GEO results back to a client: a citation event and a named mention are scored separately, because only one of them is what the buyer actually reads. The same discipline applies to earned coverage before it reaches sales: see Earned Coverage Needs a Claim-Attribution Handoff Before Sales Uses It for the five-part attribution worksheet.