Do AI Citations Actually Matter? What 354,004 AI Answers Show
Sitefire research across 354,004 AI answers on what an AI citation is worth, and why being sourced without being cited pays almost nothing.
Yes, by about a factor of 2. On days an AI model cited a brand's page, it mentioned that brand in 33% of its answers. On days the same AI model answered the same prompt and did not cite that page, 15%. That is what 354,004 answers from ChatGPT, Gemini and Google AI Mode show for 2 to 24 July 2026.
The obvious objection: brands with more coverage get cited more and mentioned more, and the citation itself does nothing. The comparison is built so that this cannot be the explanation. We never compare one brand with another. We take a prompt, ask the same AI model every day for 23 days, and look at one brand in those answers. On about 2 of those days the AI model cites the brand's page. On the other 21 it does not. The brand's coverage, size and reputation are the same on all 23 days. What changes from day to day is which pages the AI model's own web search pulled up.
Two more results follow from the same data. A page the AI model's search pulled up but did not cite is worth about nothing: 12.6% against 11.2% for a page it never pulled up. And a citation is worth most where a brand is barely mentioned today, 9.2x, and almost nothing where it already is, 1.1x.
How the comparison works
Take a prompt and ask the same AI model every day for 23 days. Pick a brand and read those 23 answers. For each answer we record 2 things: did the answer cite a page on the brand's domain, and did the answer mention the brand by name.
Then we compare the brand mention rate on cited days against not-cited days, for that brand on that prompt. We do this for every brand, prompt and AI model where both kinds of day occur, and average the differences. No model, no weighting, no adjustment.
Anything that makes one brand more visible than another, from press coverage to being the market leader, is the same on a cited day and a not-cited day, because nothing about the brand changed. Only the day did. That is what the "denser coverage" objection runs into: coverage does not change between Tuesday and Wednesday.
What does change between days is the AI model's own web search. It runs fan-out queries, pulls up a different set of pages each time, and sometimes the brand's page is among them and gets cited. The comparison measures what the mention rate looks like on those days against the others.
What a citation is worth
| AI model | not cited | cited | ratio |
|---|---|---|---|
| Gemini | 15.1% | 22.4% | 1.5x |
| ChatGPT | 15.3% | 35.5% | 2.3x |
| Google AI Mode | 15.0% | 38.5% | 2.6x |
| All three | 15.1% | 32.9% | 2.2x |
All three not-cited baselines sit within 0.3 points of each other. We did not arrange that. It comes from applying the same population rule to all three AI models, and it is one sign that they are measured the same way. The direction is the same on all three, and they are built by two different companies on different infrastructure.
It also holds brand by brand. Across the brands with enough data to measure on their own, 96% show a higher mention rate on cited days. The median brand shows 2.6x. The bottom tenth still shows 1.2x.
Now read the same 32.9% from the other end: 67% of cited pages arrive with no brand mention at all. A citation is a lever with a low hit rate. That is why one citation on one day is not a result, and why the rate at which you get cited on the prompts you care about is the number to report.
Did the AI model pick the brand first?
The alternative explanation is that the AI model decides to recommend a brand, then reaches for that brand's page to back up what it was going to say anyway. Comparing cited days against not-cited days cannot tell those two stories apart on its own.
So we split the comparison by whether the prompt itself already carries the brand's name. Where it does, the AI model has every reason to mention the brand regardless, and the cited-day gap is 4.0 points. Where the prompt does not carry it, the gap is 18.0 points.
If the pattern were mostly the AI model deciding first, those two numbers would be close. They differ by a factor of 4.5.
That constrains the alternative explanation without removing it. Retrieval and generation happen in one pass, so the citation and the mention are written together rather than one after the other. Everything in this article is a correlation measured under tight controls, not the result of changing one thing and watching another move. The only clean test is publishing pages on dates you control and watching what the AI models do, and that needs a different design.
Does being sourced pay on its own?
Most GEO reporting counts retrieval: pages crawled, pages surfaced, pages in the AI model's research. If a citation is worth a doubling, the steps leading up to one should be worth something too.
On ChatGPT we can test that, because ChatGPT is where the AI model's fan-out query is recorded. Sitefire records the pages the search pulled up (sourced) as well as the ones the answer cites (cited). That splits a brand's page into 3 states on any prompt: not sourced, sourced but not cited, or cited.
With the fan-out query held fixed, the comparison runs across answers that searched for the same thing. The brand mention rates:
| How far the page got | brand mention rate |
|---|---|
| not sourced | 11.2% |
| sourced, not cited | 12.6% |
| cited | 25.5% |
A page ChatGPT pulled up and passed over earns its brand about as many mentions as a page it never pulled up at all. The step from not sourced to sourced is 1.4 points. Run the identical comparison against the following day's exposure, where no real effect can exist, and it returns 0.2 points. The step from sourced to cited is 12.9 points.
Being sourced is not a partial win on the way to a citation. The return starts at the citation.
The floor
In answers where a brand's page was never sourced at all, the brand is still mentioned 11.2% of the time. That is 44% of the cited-state rate.
That floor is what the AI model already knows about the brand, plus the third-party pages it pulled up instead of the brand's own. It exists whether or not the brand's page is cited on the day. A citation roughly doubles a rate that does not start at zero, and a forecast built on the doubling alone will overstate what a citation buys.
Who does a citation pay most for?
Brands the AI model rarely mentions today. We grouped brands by how much AI visibility they already had on a prompt, measured on separate days from the comparison itself so that the grouping and the outcome share no observations.
| Existing AI visibility | not cited | cited | multiple |
|---|---|---|---|
| 0% | 1.1% | 10.0% | 9.2x |
| under 15% | 4.4% | 26.7% | 6.1x |
| 15-35% | 14.1% | 49.7% | 3.5x |
| 35-65% | 35.3% | 69.8% | 2.0x |
| over 65% | 77.4% | 86.6% | 1.1x |
The multiple shrinks as a brand becomes more visible, and it never reaches zero. Brands the AI model already mentions in most answers still show a positive gap, a small one.
Read the top row with its baseline: 9.2x is 1.1% going to 10.0%. It is the largest multiple in the study and the smallest number of extra mentions.
The lever is where the current numbers look worst. On prompts where the AI model already mentions you in most answers, a citation changes little. On prompts where it almost never mentions you, a citation is worth several times more.
What this does not show
This is a correlation, not an experiment. Retrieval and generation happen in one pass, so a citation and a brand mention are written together. Nothing here says what happens if you change one thing and leave everything else alone. The branded-prompt split constrains the alternative explanation without closing it.
The 3-state result is ChatGPT. ChatGPT is where the AI model's fan-out query is recorded, which is what makes the middle state visible. The 2-state result covers all three AI models.
The prompts are commercial buying-intent prompts. They describe what happens when someone is shopping, not AI usage in general.
Ratios are firmer than levels. The interval around the citation step is wide, and a forecast built on these numbers should use the range, not the point.
We cannot tell you what third-party coverage is worth to your brand. A cited page is never joined to the brands its text mentions, so whether coverage on a forum, a publisher or a comparison site helps your brand specifically is out of reach in this data.
What to do differently
- Stop counting retrieval as progress. A page surfaced and passed over performs like a page never surfaced. If your GEO reporting leads with pages crawled or pages in the AI model's research, it reports the step that does not move the outcome.
- Report citation rate on the prompts you care about, not citations in total. 2 of every 3 cited pages arrive with no brand mention, so a count of citations says little. The rate at which you get cited on one prompt says a lot.
- Aim first at the prompts where the AI models rarely mention you. That is where a citation is worth 9.2x rather than 1.1x. It is also the part of the prompt set that most reporting ignores, because the current numbers look bad.
Sitefire tracks brand mentions and citations per prompt across ChatGPT, Gemini and Google AI Mode, which is how the numbers in this article are measured. Our earlier research on why AI models cite different sources runs on the same measurement panel and covers which pages get cited. This article covers what a citation is worth once you have it.
Frequently asked questions
Do AI citations increase brand mentions?
On days an AI model cited a brand's page, it mentioned that brand in 32.9% of answers, against 15.1% on days it did not, comparing the same AI model on the same prompt about the same brand. That is a correlation measured under tight controls, not a causal result, because retrieval and generation happen in a single pass.
Does getting crawled or retrieved by ChatGPT help?
Not on its own. A page ChatGPT sourced but did not cite earns its brand about as many mentions as a page it never sourced, 12.6% against 11.2%. The measurable return begins at the citation.
Which AI model rewards citations most?
Google AI Mode showed the largest multiple at 2.6x, ChatGPT 2.3x and Gemini 1.5x. All three started from almost the same baseline, within 0.3 points of each other.
If I am invisible in AI answers, is GEO worth it?
That is where the multiple is largest. For brands an AI model rarely mentioned otherwise, the cited state showed 9.2x the brand mention rate of the not-cited state, against 1.1x for brands already mentioned in most answers.
Does one citation fix a prompt?
No. 67% of cited pages arrive with no brand mention at all. A citation is a lever with a low hit rate, which is why the rate matters more than any single instance.
Key takeaways
- A citation is worth about a doubling of a brand's AI visibility: 15.1% to 32.9%, positive on all three AI models and for 96% of brands measured.
- Being sourced without being cited is worth almost nothing. That step moves the brand mention rate 1.4 points, against 0.2 for a placebo.
- The floor matters. Brands are mentioned 11.2% of the time with no page sourced at all, which is 44% of the cited-state rate.
- 2 of every 3 cited pages carry no brand mention, so citation rate on specific prompts beats citation counts.
- The multiple is largest where a brand is currently least visible, 9.2x against 1.1x.
Methodology
Two Sitefire analyses on one measurement panel, with the data fixed before analysis.
The first covers 354,004 answers from ChatGPT, Gemini and Google AI Mode between 2 and 24 July 2026. The unit is one brand on one prompt, as answered by one AI model across the window. Within each unit we compare the brand mention rate on days that brand's page was cited against days it was not, then average across units with no weighting. A unit contributes only if both kinds of day occur.
The second covers the 132,938 ChatGPT answers inside that same window. ChatGPT is where the AI model's fan-out query is recorded, which is what makes the middle state visible. Holding that query fixed, a page is classified as not sourced, sourced but not cited, or cited, and the 3 rates are computed on the same set of prompt and brand pairs so the levels chain.
The two analyses report the doubling from different starting points. The first compares cited against not cited across all three AI models and reports 15.1% to 32.9%. The second compares cited against sourced-but-not-cited on ChatGPT with the fan-out query held fixed, and reports 12.6% to 25.5%. Both land near a factor of 2. The difference is a population difference, not a disagreement.
A brand mention is the brand's name appearing in what the answer says. Where an AI model renders a citation using the site's domain as the visible link text, that occurrence is excluded, since otherwise the citation and the mention would be produced by the same characters.
Intervals are computed by resampling whole companies rather than individual observations, because observations within one company are not independent. We publish the answer count and the AI models; other counts stay internal.
Sources
Sitefire proprietary analysis. 354,004 AI answers across ChatGPT, Gemini and Google AI Mode, 2 to 24 July 2026, including 132,938 ChatGPT answers for the three-state analysis.
