Key takeaways
- A mention is a yes/no. Position, frequency, and framing tell you whether the mention actually helps you.
- ChatGPT answers are non-deterministic, so one check is a single draw. Run each prompt many times and track averages.
- Almost half of ChatGPT responses cite 10 or more domains, so you are always sharing the answer with competitors.
- Separate three things: mentioned, cited, and recommended. They fail for different reasons and need different fixes.
- You can start with a spreadsheet for a month. Move to a tracker when the manual work stops scaling.
Why a mention is not enough
You open ChatGPT, ask "what are the best tools for X," and your brand shows up. Nice. Ask again two hours later and it might not. Ask a slightly different version and you are fourth in a list of six, described in one flat sentence, while a competitor gets a paragraph.
Those are all "mentioned." They are not the same outcome.
Position matters because people read answers top-down and remember the first names. Some brand monitoring roundups list "recommendation position" as a core metric for this reason: a brand named first in a list-style answer tends to get more attention than one named later. That is plausible, though I would treat it as a working assumption, not a law. What you can measure for certain is where you land, how often, and with what wording.
One warning before the method. Be careful with the phrase "we're third in ChatGPT." There is no permanent ranking. There is a distribution of positions across many runs. Report that distribution, not a single number.
The metrics that describe position
Tools name these differently, but the useful set is small.
| Metric | What it answers | What it does not prove |
|---|---|---|
| Mention rate | In what share of runs does your brand appear at all? | That you are recommended or well described |
| Average position | Where in the answer do you appear (first named, third, last)? | That position drives clicks |
| Visibility score | A blended score of placement, share of attention, and depth | Causes. It is a summary, not a diagnosis |
| Citation rate | How often is your own URL listed as a source? | That the brand is named in the text |
| Share of voice | Your mentions as a share of all brands named | Quality of the mentions |
| Sentiment | How the answer describes you | Whether the description is accurate |
The distinction between mentioned and cited is the one people skip. ChatGPT can pull your page as a source and never name your brand in the text. It can also name you from training data with no link to your site. Surfer Academy's walkthrough on reverse-engineering brand visibility in ChatGPT calls out the "cited but not mentioned" case specifically, and it is worth treating as its own bucket.
Here is a quick way to read the combination:
- Low mention rate across many prompts: a coverage problem. You are not in the consideration set.
- Mentioned, but late in the answer: a prominence problem. You are known but not preferred.
- Mentioned without your own citations: a content or crawlability problem. The model knows you from elsewhere and does not use your pages.
- Cited but not named: your content is feeding answers that credit someone else, or the page does not tie the facts to your brand name.
- Mentioned with weak wording: a reputation problem, usually driven by third-party sources.
Step 1: Build a prompt set from real demand
Write prompts a buyer would type without knowing your name. Questions, not keywords. Practitioners on Quora and Reddit suggest anywhere from 10 to 15 prompts at the low end to 20 to 40 for a serious baseline. I would start with 25 and cluster them.
Useful clusters:
- Organic category prompts: "best project management software for small teams"
- Comparison prompts: "Asana vs Monday.com"
- Brand-specific prompts: "is [your brand] good for small teams"
- Use-case and "alternatives to" prompts
The clusters matter because ChatGPT behaves differently in each. Per Promptwatch Data (Sep 28 to Oct 4, 2026), 52.5% of organic prompts cite 10 or more domains, while for brand-specific prompts the most common count is 5 domains and only 21.3% reach 10 or more. Your brand-specific prompts are a narrower arena. Your organic prompts are the crowded one where position is hardest to win.
A good source of prompts is your own Google Search Console queries. Reword the question-shaped ones into natural prompts. If you use Google Search Console, [tool:google-search-console] is free and already has the data.
Step 2: Run each prompt many times
This is the step that separates measurement from anecdote. ChatGPT answers vary run to run. One practitioner reported that two answers to an identical prompt shared under a tenth of their sources. A single check is a single draw from that spread.
Rules I would follow:
- Run every prompt at least 5 to 10 times per measurement period.
- Use a logged-out or clean session. If your account has discussed your company for months, the model has context you are not accounting for and you will overestimate your visibility.
- Record the model version and keep the raw response text. Once you change your site, the "before" is gone.
- Do your baseline before shipping any content changes. Otherwise every claim about what worked is a story, not a measurement.
- Use the same cadence every time. Weekly is fine to start.
Step 3: Log position, not just presence
For each response, capture a row. Here is a simple schema for a spreadsheet:
prompt | date | model | run # | mentioned (Y/N) | position (1st, 2nd... or N/A) | brands named total | cited URL (Y/N) | descriptor text | competitors named before you | sources listed
A few notes on defining position, since I found no universal standard in the sources I read:
- For list-style answers, use the ordinal slot in the list.
- For prose answers, use the order in which brands first appear.
- Also record how much text each brand gets. A one-line mention in slot two can be weaker than a full paragraph in slot four.
- Note when the answer opens with a single recommended pick. That is a different outcome from a neutral list.
Then compute, per prompt and per cluster: mention rate, average position when mentioned, and the share of runs where you are in the top 3. Trend those over time. Ellevate's guidance on this is sensible: track frequency of appearance, and use monitoring to see whether your presence is changing, not to claim a fixed rank.
Step 4: Capture the context around the position
Position without context is half the picture. Three things to log alongside it.
Who you appear next to
Promptwatch's research says to plan on sharing every answer, and to check which competitors are cited in the same responses as you. If the same two rivals sit above you in every list, that is your actual competitive set in AI answers, which may differ from the one in your sales deck.
How you are described
Copy the exact sentence about you. Look for stale facts, wrong pricing, or generic wording like "a tool for teams." Descriptors are often where the real leverage is, and they usually trace back to a handful of third-party pages.
What sources sit behind the answer
Log citations as carefully as mentions. Promptwatch's content-type data (Sep 28 to Oct 4, 2026) shows product pages are the most cited type at 22.4%, followed by landing pages at 16.8% and listicles at 13.6%, and for commercial prompts product pages rise to 26.0%. So if you are absent, check whether your product and landing pages are even eligible, and whether listicles that rank you are placing you low.
Query fan-outs are relevant here too. ChatGPT runs its own web searches behind an answer, and Promptwatch's fan-out data suggests short, query-like headings with the entity and category up front help your pages match those searches. The data also notes the average number of searches per response moved around over time, so treat any single figure with caution. On August 8, 2026, ChatGPT began using the site: operator at scale, which jumped from about 0.4% to about 17% of fan-out queries. Sudden shifts like that can move your position without any change on your side, which is another reason to keep a dated log.
Step 5: Do it by hand for a month, then decide on tooling
I agree with the advice that surfaced in the research: do it manually for a month first. You will learn what you are actually buying, and you may find the manual version is enough for a small prompt set.
It stops being enough when you have more than about 30 prompts, several markets, or multiple models. At 30 prompts times 8 runs times 3 platforms, you are logging 720 responses a cycle.
Tools that track position and not just mentions
Not every tracker reports position. Some only report presence or share of voice. When evaluating, ask three questions: does it run real user interfaces or only the API, does it repeat prompts to give you averages, and does it log position and cited sources per response.
Promptwatch tracks real ChatGPT, AI Overviews, AI Mode, Gemini, Perplexity and other interfaces and reports Visibility Score, Share of Voice, position in the answer, sentiment, and citation rank. It also goes past monitoring: IP-verified AI crawler logs show whether ChatGPT's bots are reading your pages, and its content gap analysis and Content Agent help close the gaps you find. That matters if your diagnosis is "mentioned but not cited," because the fix lives on your site.
Other trackers worth a look, depending on budget and need:
| Tool | Tracks position in the answer | Citation and source data | Goes beyond monitoring | Best for |
|---|---|---|---|---|
| Promptwatch | Yes, with Visibility Score and position | Yes, plus crawler logs and AI traffic | Yes: content gaps, Content Agent, Action Board | Teams that want to measure and fix |
| Otterly.AI | Presence and share-style metrics | Basic | No | Low-cost monitoring |
| Peec AI | Visibility and position metrics | Yes | No | Straightforward tracking |
| Profound | Yes | Yes | Limited | Enterprise monitoring budgets |
| Keyword.com | Presence, prompt clusters, competitor positioning | Basic | No | SEO teams already using its rank tracker |
| LLMrefs | Rankings across models | Yes | No | Keyword-style tracking |
I have not verified every feature above hands-on, so confirm current plans on each vendor's pricing page. For a broader directory, the GEO software directory at bestgeosoftware.com lists more options.
Common mistakes
- Reporting one run. It tells you almost nothing.
- Testing while logged into an account that knows your brand.
- Treating all prompts as equal. Brand-specific prompts flatter you. Organic ones are the honest test.
- Ignoring citations. A mention with no self-citation is fragile.
- Tracking only ChatGPT. A brand can be strong on one platform and nearly absent on another, so prioritize by where your buyers are.
- Changing the prompt set midstream. Add new prompts, but keep a stable core so trends stay comparable.
What to do when your position is weak
Once you know which problem you have, the next moves are concrete.
- Late position: earn placement in the third-party listicles, reviews, and forum threads that ChatGPT already cites for your category. Promptwatch's research recommends earning mentions on third-party reviews, listicles and forums.
- No self-citations: check crawl access and make sure product and landing pages state your category, audience, and differentiators plainly.
- Weak descriptors: fix the source pages that describe you, and update your own pages so the facts are easy to quote.
- Missing from comparisons: publish honest comparison and alternatives pages with specifics.
Re-measure after each change using the same prompts and run count. If the average position does not move after a few cycles, your hypothesis was wrong, and that is useful too.
If you would rather hand the technical SEO and content work to specialists, agencies like 1001 SEO Media do this kind of AI search work as part of their GEO service.
A starter checklist
- Write 25 buyer-style prompts in four clusters.
- Run each 5 to 10 times in a clean session, weekly.
- Log mention, position, descriptor, competitors above you, and cited sources.
- Compute mention rate, average position, and top-3 share per cluster.
- Diagnose: coverage, prominence, crawlability, or reputation.
- Pick one fix, ship it, and re-measure on the same set.
Position inside the answer is a noisy number. Treat it as a trend across many runs and it becomes the most useful signal you have about whether ChatGPT is actually recommending you.