Key takeaways
- Traditional rank tracking returns a deterministic number for a keyword, location and device. ChatGPT "position" is probabilistic: the same prompt can give a different order, different citations, or no mention of you at all on the next run.
- A typical ChatGPT answer is a shared list, not a ladder. In Promptwatch's data for Sep 28 to Oct 4, 2026, 47.6% of cited responses named 10 or more domains.
- Measure presence, position inside the answer, and how you are described, aggregated across many prompts and many runs. A single snapshot tells you very little.
- Under the prompts sits a retrieval layer of fanout queries that can change overnight, as it did on August 8, 2026. Keep your classic rank tracker for Google, and add a tracker built for sampling.
Why a rank number used to be enough
For twenty years, rank tracking was simple. You gave a tool a keyword, a country and a device. It asked Google, read the result page, and told you your URL sat at position 4. Run it again an hour later and you got the same answer, give or take a shuffle. That determinism is why the whole industry built reporting, bonuses and client retainers around it.
ChatGPT does not work like that. There is no results page with ten slots. There is a generated answer, assembled from web searches the model decides to run, written in prose, and different every time you ask. "Where do we rank in ChatGPT?" sounds like the same question as "where do we rank in Google?" It isn't, and tools that pretend otherwise will give you tidy numbers that mean less than they look.
Traditional tracking still matters. A client can hold position one for a target keyword and still be invisible to the buyer who asked ChatGPT to recommend a vendor in the category. Both facts are true at once, and you need to see both.
How the two kinds of tracking differ
| Dimension | Traditional rank tracking | AI search tracking |
|---|---|---|
| Unit of measurement | Keyword | Prompt (a full conversational question) |
| Output | A position, 1 to 100 | Mention, citation, position in the answer, sentiment |
| Repeatability | High: same input, nearly same output | Low: varies by run, model version, locale and phrasing |
| What you rank | A URL | A brand, a product, or a cited source |
| Sampling needed | One check per day is fine | Multiple runs per prompt, aggregated over many prompts |
| Click assumption | Higher position means more clicks | Often no click at all; the mention is the impression |
| Underlying mechanism | Index and ranking algorithm | Model knowledge plus live retrieval via fanout searches |
One way to say it: traditional trackers ask where your page sits. AI trackers ask whether you get named, how early, and what the model says about you.
What "position" means inside a ChatGPT answer
There are really four different things people mean when they say position in ChatGPT.
Position in the generated text
If ChatGPT lists five tools in response to "best project management software," your brand might be first, third, or missing. This is the closest analogue to a rank number, and it is the one tools usually report. It is also the least stable, because the model can reorder the same candidates between runs.
Position in the source list
ChatGPT shows citations alongside the answer. Where your domain lands in that list is a separate signal from where your brand lands in the prose. You can be cited and never named, or named and never cited. If you are mentioned without self-citations, that points to a content or crawlability problem rather than a reputation problem.
Share of the answer
This is the awkward one for people used to rankings. Per Promptwatch Data, 47.6% of ChatGPT responses with at least one citation cited 10 or more unique domains between Sep 28 and Oct 4, 2026, and 66.1% cited at least 8. Only 6% cited three or fewer. For general questions that name no brand, 52.5% cited 10 or more domains.
So a typical answer is a crowd. You are not fighting for the one top slot, you are trying to be one of the ten sources in the room, and ideally one of the first few brands named.
Brand-specific prompts behave differently. When the prompt names one brand, the most common count is 5 domains (14.9% of responses), and only 21.3% cite 10 or more. If someone asks about you by name, the answer is narrower, and who else gets cited alongside you matters a lot.
Position across the distribution
The only honest "rank" is a distribution: across 100 runs of this prompt set, you appeared in 62% and were named in the first two positions in 24%. That is a different kind of number than "4," and it is the kind a decent tool should give you.
Why the same prompt gives different answers
The research behind most vendor guides agrees on the core point: ChatGPT visibility is probabilistic. Model temperature and stochastic generation mean the same prompt can produce different brand orderings, citations, or omissions across runs. On top of that:
- Different model versions recommend different brands for the same question.
- Region and language change the answer.
- Small wording differences ("best CRM" versus "best CRM for a 10-person agency") change the candidate set.
That is why checking ten prompts by hand and pasting results into a spreadsheet falls apart quickly. It isn't a measurement, it is an anecdote with a timestamp.
The layer most people miss: query fanouts
Here is the part that surprised me when I looked at the data. ChatGPT doesn't search the web with the user's literal prompt. It generates its own search queries, called fanouts, and builds the answer from those results.
Promptwatch's ChatGPT query fanout report shows how much that layer moves. Average fanouts per response were 2.15 in early December and about 1.84 by early March, and average query length fell from roughly 117 characters to the high 80s, then to about 53 characters in April. The report's own text says a prompt can trigger "3-8+" searches, which is out of step with those averages, so treat the headline range with some caution and look at the trend lines.
Then there is August 8, 2026. Promptwatch's analysis of the site: operator found that the share of fanout queries containing site: jumped from 0.37% to 16.8% in a single day, and average fanouts per response rose from about 1.08 to about 1.83. Nobody on the brand side changed anything. The platform did.
Two consequences follow:
- Any keyword-to-fanout mapping you built before August 8 is stale.
- Your own domain now acts more like a retrieval surface. Crawlability and indexing gaps turn into gaps in AI answers.
A classic rank tracker has no way to see any of this. If your visibility dropped overnight and your Google positions didn't move, the cause may sit entirely in how ChatGPT searches.
What to measure instead of a rank number
If a single position is the wrong unit, what replaces it? A short list that works in practice:
- Visibility rate: the share of tracked prompts, across repeated runs, where you are mentioned or cited.
- Position in the answer: average and distribution, not a single snapshot.
- Share of voice: your mentions against competitors across the same prompt set.
- Citation rank and self-citation rate: whether your own pages are the sources.
- Sentiment: how the answer describes you, since a mention that calls you "a budget option with weak support" is not a win.
- Co-citation: which competitors show up in the same responses as you.
- AI referral traffic and conversions, so visibility connects to revenue.
Read together, they point at the problem. Low share of voice is a coverage problem. Late position is a prominence problem. Mentions without self-citations suggest content or crawler access trouble. Weak sentiment is a reputation problem. Promptwatch's own advice from the domain-count report is blunt: plan on sharing every answer, track where you appear and how ChatGPT describes you, measure across many prompts, and earn mentions on third-party reviews, listicles and forums.
The buyer's guide above is a useful read on procurement questions. Its core argument matches the data: reliable trackers sample each prompt several times a day, keep history, and aggregate across hundreds of related prompts, so noise at the single-prompt level becomes a trend at scale.
How to run both systems side by side
You don't have to choose. They answer different questions.
Keep the classic tracker for Google
Organic blue links still drive a large share of traffic, and you need a baseline, an early warning for drops, and local tracking where it applies. Established rank trackers are fine here.
AccuRanker and Advanced Web Ranking both added AI search monitoring on top of their SERP tracking, which is convenient if you want one dashboard. Just check how they sample AI answers before trusting the numbers.
Add a tracker built for sampling
The criteria I would insist on:
- Real front-end monitoring rather than API-only. User-facing answers and citations can differ from API outputs.
- Repeated runs per prompt, with distributions and history.
- Country and language control.
- Exact cited URLs, not just brand mentions.
- Some way to act on what you find.
Promptwatch covers that list and goes further than most. It monitors the actual interfaces of ChatGPT, Google AI Overviews, AI Mode, Gemini, Perplexity and others, reports Visibility Score, position, share of voice and sentiment, and pairs that with IP-verified AI crawler logs and visitor analytics so you can see whether a visibility change connects to crawling or traffic. It also shows fanout queries and prompt volumes, which matters given the August shift. Its Action Board and Content Agent then turn gaps into published content, which most monitoring-only tools don't attempt.
Other options worth a look, depending on budget and team:
For a wider browse, the GEO software directory at bestgeosoftware.com and the AI rank tracking directory list far more.
Quick comparison by tool type
| Tool type | Good at | Weak at | Examples |
|---|---|---|---|
| Classic SERP rank trackers | Google positions, local, SERP features | Prompts, citations, sentiment | AccuRanker, Advanced Web Ranking |
| SEO suites with AI add-ons | One login for SEO and light AI tracking | Often fixed prompt sets, shallow AI data | Semrush, Ahrefs Brand Radar |
| Prompt trackers | Cheap monitoring of mentions and position | Crawler data, traffic, content action | Peec AI, Otterly.AI |
| AI visibility platforms | Tracking plus crawler logs, traffic and content workflows | More to learn, higher price than basic trackers | Promptwatch, Profound |
A practical weekly rhythm
- Build a prompt set of 50 to 150 prompts covering general questions, brand-specific questions, and competitor comparisons. Don't skip the comparison prompts; they are where buying decisions get shaped.
- Track by country and language where you actually sell.
- Review the distribution, not single answers. Look at visibility rate and average position week over week.
- Check which pages get cited and which competitors appear next to you.
- Look at crawler logs and site health when visibility dips. Often it is access, not content.
- Compare against Google data. Pages that rank in Google but never get cited by AI are your best content-gap candidates.
- Re-baseline when the platform changes. If fanout behavior shifts, old assumptions go out the window.
What to tell your boss or client
The awkward conversation is that you can no longer promise a position. You can promise a trend in visibility rate across a defined prompt set, a clearer picture of who is cited next to you, and a link from those mentions to traffic. That is a better deal than a rank number, honestly, because a rank number in ChatGPT would be fiction.
Some vendor claims about AI referral traffic converting at several times the rate of organic search circulate widely. One agency source puts it at 4.4x, but I couldn't verify it, so test it on your own analytics before repeating it.
Bottom line
Position in Google is a place on a page. Position in ChatGPT is a likelihood of being named, somewhere inside an answer that cites around ten domains, produced by searches the model writes itself. Keep tracking Google the old way, and add sampling-based tracking for AI answers. Then judge both by what actually reaches your site. If you want a deeper dive into which AI answers cite what, Promptwatch's data hub is a good place to start.