Suppose you ask ChatGPT to recommend SaaS consultants in Taiwan, see your company in the answer, and breathe a sigh of relief. That is a lucky draw, not an audit. Change accounts, disable memory, or ask again tomorrow and the answer may change. A useful audit measures three things under controlled conditions: how often the brand appears, where it ranks in the answer, and whether the answer includes a clickable source.
Separate the three levels of citation
Teams often combine three distinct outcomes and end up with a misleading result. An answer may mention the brand without a link, cite the website in a source card, or recommend the brand as its first choice. Those outcomes have very different commercial value, so record them separately.
- Unlinked mention: The brand name appears, but there is no link or meaningful recommendation order. It establishes a baseline, yet competitors can easily crowd it out.
- Source citation: The answer includes a card or link to your URL. Users can visit the site, making this the outcome most GEO programs pursue.
- Top recommendation: When asked to choose one provider, the model names your brand first or devotes the answer to it. This is the rarest and most valuable signal.
Control four variables before you test
Treat the audit like an experiment. If you test through a signed-in account today and an incognito window tomorrow, personalized memory can change the answers and make the results incomparable. Define the following four conditions before you begin.
- Account and memory: Stay logged out or use an incognito window, then disable ChatGPT's Reference Chat History to remove personalization from the test.
- Platform and mode: Test ChatGPT, including Search mode, Perplexity, Google AI Overviews, and Gemini separately. Never combine their results.
- Language and region: Use the language your target customers use. For Taiwan B2B buyers, test in Traditional Chinese; English answers do not represent the same market.
- Timing: Run the full prompt set on the same day whenever possible. Model indexes and live retrieval change continuously, so results collected weeks apart are not comparable.
Five prompt templates you can copy
One question cannot show the full picture. Buyers approach the same need through category searches, named comparisons, and specific situations. The five prompt types below cover the journey from first discovery to active consideration. Replace [category], [brand], and [situation] with your own terms.
- Category discovery: "I need a [category, such as a B2B GEO/SEO agency in Taiwan]. Recommend several options and explain what each does well." Check whether you appear on an unprompted shortlist.
- Named comparison: "How does [your brand] compare with [main competitor] for [service]?" Check whether the model recognizes your brand and describes it accurately.
- Situational need: "I run a SaaS company in [industry] and want more visibility in AI search. Who should I contact, and where should I start?" Check whether the answer includes you in the solution.
- Fact check: "What services does [your brand] provide, and what are its prices and service scope?" Check whether the answer is accurate and current.
- Source request: "Recommend [category] providers and include the source URLs you used." This forces the answer to reveal its sources so you can see whether it links to your site.
Measure ChatGPT and Perplexity separately
The two systems handle citations differently. Perplexity attaches numbered sources to nearly every answer, so you can count where your URL appears and how often it is cited. With ChatGPT, separate answers produced from model memory from answers produced with Search. Long-term brand visibility shapes the former; live web retrieval shapes the latter. Test each prompt in both modes, because the difference between them is useful evidence.
Turn the answers into one scorecard
Do not judge the results from memory. Create a table with four platforms across the columns and five prompt types down the rows, then record three values in each cell. Repeat the same table later and you will have a trend instead of a one-off impression.
- Mention (0 or 1): Did the answer name your brand?
- Citation level (1 to 3): Score an unlinked mention as 1, a source link as 2, and a top recommendation as 3.
- Competitor coverage: Note which competitors appear in the same answer and whether they rank above or below you.
Use the audit gap to choose the fix
An audit should lead to a specific repair, not a vague instruction to publish more. If the brand receives no unlinked mentions, it probably lacks third-party discussion and corroboration. If it earns mentions but few source citations, its pages may be difficult to extract: perhaps the question-and-answer structure is unclear, schema is missing, or key facts sit inside images and long paragraphs. The first is a visibility problem; the second is a technical one.
You can run this audit in about half a day and learn whether the main gap is visibility or structure. Tenten GEO's 30-day GEO audit takes the process further by organizing the twenty-cell scorecard, competitor comparison, and repair plan. To see the gap before committing to a full audit, book a 30-minute GEO diagnostic session and we will test your brand with you.
