Practical review · GEO · 16 Oct 2026
Do ChatGPT, Perplexity and Google AI cite your brand? Why a single measurement proves nothing and how to build a sample you can trust
A single citation check in ChatGPT or Google AI proves nothing. It is more reliable to count the share of answers where the brand appeared across many runs, with the conditions recorded.
- Same brand listrepeat of one prompt<1%
- Same list in the same order<0.1%
- Bose, Sony, Sennheiser, Apple in answersheadphone prompt, 994 answers55–77%
SparkToro study as relayed by Search Engine Journal, 30 Jan 2026. The study page and raw data were not opened.
In short
- In the SparkToro study (2,961 runs, November–December) ChatGPT and Google AI returned an identical brand list in under 1% of repeats, and the same list in the same order in under 0.1%.
- It is more reliable to count the share of answers where the brand appeared across many runs than its place in a list.
- The cited article does not provide raw data for an independent recalculation; treat its result as a finding from one sample.
Why a single measurement proves nothing
What the SparkToro study found (as relayed by Search Engine Journal):
- 12 brand prompts (knives, headphones, hospitals, consultants and more), each run 60–100 times per platform, 2,961 runs in total;
- Claude repeated the list slightly more often but repeated the order less often; no platform proved reliably repeatable;
- 142 participants wrote their own headphone prompt; the wording barely overlapped (similarity score 0.081), yet Bose, Sony, Sennheiser and Apple appeared in 55–77% of 994 answers;
- brand frequency was more stable than brand position in the list; in narrow categories leaders showed more clearly than in broad ones.
Open questions the authors did not answer: how many runs are needed for reliable data, and whether the API behaves like a manual query.
How to build a sample of prompts and repeats
This is the guide’s working scheme, not a proven norm: the study did not determine how many repeats are enough.
- Take 20–50 prompts along the customer’s buying path, with 2–3 phrasings each.
- Run each prompt many times. The study used 60–100 runs; a practical minimum for a first estimate is 10–20, then scale up.
- Add competitors to the same prompts so the brand’s share has something to be compared against.
- Record the conditions of every run.
| What to record | Why |
|---|---|
| Platform and mode (ChatGPT, Perplexity, AI Overviews, AI Mode) | answers of different systems cannot be mixed |
| Date and time | answers change over time |
| Region and language | results depend on place |
| Logged in or not | personalisation distorts the measurement |
| Exact prompt wording | for reproducibility |
| Brand mentioned / link to your domain present | a mention and a link are different results |
Does AI cite the brand: a six-question check
Answer “yes”, “no” or “don’t know” about your measurement.
- Was each prompt run many times rather than once?At least 10–20 repeats
- Did the wording vary rather than stay single?2–3 phrasings per prompt
- Are platform, mode, region, date and login state recorded?Conditions of each run
- Do you count the share of answers with the brand, not list position?Frequency is more stable than order
- Are competitors measured on the same prompts?Something to compare against
- Do you distinguish a mention from a link to your domain?Two different results
This guide’s own working scale, not an industry standard. “Don’t know” counts as “no”.
How to record and present the result
- Phrase it like this: “the brand appeared in 14 of 40 answers (35%), measurement dates, platform, number of repeats”. The numbers here illustrate the format, they are not data.
- Show mentions and links to your domain separately.
- Do not promise a place in the answer: the study found order is unstable.
- Cross-check against an independent signal: impressions in the Search Console Generative AI report (see the AI impressions guide). They count only shown links.
- Repeat the measurement on a schedule and compare periods on the same prompts.
FAQ
Can paid trackers be trusted?
They also measure a sample. Ask how many runs per prompt, how repeats work and what exactly counts as a brand appearance.
Do I need to vary the wording?
Yes: in the study different wordings led to a similar set of brands, so one wording gives a narrow slice.
Do a manual query and the API give the same result?
The study authors did not establish this.
How many repeats are needed?
The source does not say. The guide’s working minimum is 10–20 per prompt for a first estimate.
Sources and method
- Search Engine Journal: AI recommendations change with nearly every query (SparkToro) — retelling of the study, 30 Jan 2026
Method: the SparkToro study page and its raw data were not opened; figures come from the retelling. The working minimum of repeats and the scale are guide proposals. Data as of 16 Oct 2026.