Key takeaways
- Single-run "rank" in ChatGPT, raw mention counts, generic AI visibility scores, cross-engine citation counts, and domain authority as an AI proxy are the five metrics most worth retiring.
- Each engine cites a different number of sources per answer, so one blended citation number compares things that can't be compared.
- Citations decay. A one-off "were we cited this month" snapshot hides whether you are keeping or losing ground.
- Replace vanity numbers with a short set: share of voice per prompt cluster, citation persistence by content type, AI referral sessions, and conversions.
- Report the question each metric answers. If a number can't change a decision, it doesn't belong in the monthly deck.
Why this list exists
I keep seeing the same slide in AI visibility decks. A big green number, usually called something like "AI Visibility Score: 72," with an arrow pointing up. Nobody in the room can say what would have to happen for it to point down, or what they would do about it.
That is the test I use for every metric below. Does it change a decision? If not, it is decoration. And decoration gets expensive when the CMO starts asking why the green number didn't produce pipeline.
The measurement problem is real, and it is not just marketing's fault. Supermetrics argues that you can still measure AI search, but only by combining a handful of numbers from different places and pulling them on a disciplined schedule. Their four: AI impressions, citation share, AI referral sessions, and prompt-level data. That's a good starting point, and a much shorter list than most dashboards offer.
Here is what I would drop, and why.
Metric 1: Single-run "rank" in ChatGPT
This is the most tempting one to retire first. A tool runs a prompt once, finds you in the third position of the answer, and reports "ranked #3 in ChatGPT."
The problem is that AI answers are not a results page. They are generated, they vary run to run, and the searches behind them change. Promptwatch's ChatGPT query fanout data shows average fanouts per response moving from 2.15 in early December 2025 to about 1.84 by early March 2026, then 1.0 in April, with average query length dropping from roughly 117 characters to about 53. The queries ChatGPT actually runs are a moving target, and your prompt is only the starting point.
There is another shift. On August 8, 2026, ChatGPT started using the site: operator at scale, jumping from about 0.4% to about 17% of fanout queries overnight, according to Promptwatch Data. If retrieval behavior can change that fast, a single position reading is noise.
What to report instead
Report a rate across many runs: how often you appear for a cluster of prompts, with position as a secondary band (early, middle, late), not a rank. Share of voice against named competitors is the honest version of this number.
Metric 2: Raw mention counts
"We were mentioned 1,400 times this month." Great. In what context? Next to which competitors? Were you the recommendation or the cautionary tale?
A mention count rewards volume and ignores quality. A brand can be mentioned constantly in answers like "avoid X because of pricing complaints" and the count still goes up. It also ignores whether the engine linked to you. A mention with no self-citation usually points to a content or crawlability problem, not a win.
If you must report mentions, pair them with sentiment and with whether your own domain was cited. Better still, report the three together as a single row per prompt cluster.
Metric 3: Generic "AI visibility scores"
I have mixed feelings here. Composite scores are useful inside a single tool, because they let you watch a trend. They are bad in an executive deck because nobody outside the tool knows the formula, and two vendors will give the same brand wildly different numbers.
Worse, a composite hides the diagnosis. A low score could mean you aren't covered at all, you show up late, or you are mentioned without being cited. Those need different fixes: coverage is a content gap, prominence is an authority problem, and missing citations are often technical.
If your platform gives you a score, keep it for the working team. In the leadership report, show the components it is built from.
Metric 4: One blended citation count or citation rate
This one catches smart people. You add up citations across ChatGPT, Google AI Overviews, Perplexity, and Copilot, and report a single number or rate.
The engines don't have the same number of slots. Promptwatch's Average Sources Per Response report shows ChatGPT citing about 5 sources per web-search answer, Google AI Overviews about 10, and Perplexity almost exactly 10 every day. Copilot's average swung from under 2 to nearly 17 within a few weeks before settling at the low end.
So a blended figure is mostly a measure of which engine you happened to sample. And a Copilot swing is more likely to be a retrieval change on Microsoft's side than anything you did. Promptwatch's advice is to judge Copilot on monthly trends, not weekly snapshots.
Model updates move the baseline too. Promptwatch's citation drop report shows average citations per ChatGPT response falling across GPT-5.3, GPT-5.4, and GPT-5-Mini around the March 4, 2026 rollout. If you reported a "drop in citations" that week, it was probably the whole platform and not you.
What to report instead
Report citation share per engine, side by side, never summed. Perplexity is steady enough to double as a test bench: if your pickup there changes after you change a page, that is your change.
Metric 5: Domain authority as a proxy for AI visibility
Some teams still say, "Our DR is 80, so we should be fine in AI search." The data doesn't back that up. In Promptwatch's August 2026 breakdown by domain rank, sites with DR 46-75 earned almost half of ChatGPT citations, while DR 91-100 fell to about 3%.
Authority still matters for classic rankings. It just isn't a stand-in for whether a model picks your page for a specific question. Content type matters more than most people expect: in Promptwatch's content type report, product pages were the most cited type at 22.4% of ChatGPT citations (Sep 28 to Oct 4, 2026), ahead of landing pages and listicles.
Keep authority metrics for link building. Don't put them in the AI section of the report.
Metric 6: The one-off monthly citation snapshot
"We were cited on 340 pages this month." It sounds concrete. It hides the thing that matters, which is how long those citations last.
Promptwatch's content half-life report tracked every page ChatGPT cited on July 6, 2026 for 90 days. After 30 days, ChatGPT still cited 78% of landing pages and 76% of docs. By day 60 that was 69% and 67%. Product pages, reviews, and how-tos sat around 55% at day 60, with half-lives of 74 to 77 days. News was the outlier: half of news pages were gone after 36 days, and almost a quarter were never cited again after day 0.
That means a newsy content program can post a huge citation count one month and a hole the next, while a docs-and-landing-page program looks dull and stays put. A snapshot can't tell those apart.
What to report instead
Citation persistence by content type: of the pages cited 30 and 60 days ago, how many still are. It also tells you where to spend. Content that decays fast needs a refresh schedule.
Metric 7: Impressions with an invented click-through rate
Google announced Search Console's generative AI performance reports in June 2026, covering impressions for AI Overviews, AI Mode, and AI features in Discover by page, country, device, and date. Useful. But Supermetrics notes the reports carry no click, CTR, or query data, so you can't turn AI impressions into a CTR.
Some decks try anyway, dividing referral sessions by AI impressions and calling it CTR. That mixes two different data sources with different definitions. Report AI impressions as exposure, and AI referral sessions as traffic, and leave the ratio alone.
Bing is a little further along. Supermetrics notes Bing Webmaster Tools launched an AI Performance report in public preview in February 2026 and later added Citation Share, which makes it worth a look as a first-party source.
Metric 8: Blended social citation numbers
Teams often report "community and social citations" as one line. The engines behave very differently. From Sep 28 to Oct 4, 2026, ChatGPT got 0.46% of its citations from social platforms, while Google AI Overviews got 12.2%, per Promptwatch Data. And reddit.com's share of ChatGPT citations fell from about 4% to 0.5% on August 14, 2026, per Promptwatch's Reddit report.
If you told leadership last spring that Reddit was your AI strategy, that line has to be reported per engine now. A blended number would show a decline in one place and hide a stable channel in another.
What to report instead: a short scorecard
Here is the replacement set. It is deliberately small.
| Retire | Why it misleads | Replace with |
|---|---|---|
| Single-run ChatGPT rank | Answers and fanouts vary run to run | Appearance rate and share of voice across prompt clusters |
| Raw mention count | Ignores context, sentiment, and links | Mentions with sentiment and self-citation rate |
| Generic visibility score | Hides the diagnosis | The components: coverage, position, citation |
| Blended citation count | Engines cite 5, 10, 10, and a volatile number of sources | Citation share per engine, never summed |
| Domain authority proxy | DR 46-75 earned almost half of ChatGPT citations | Citation share by content type and page |
| Monthly citation snapshot | Citations decay at different speeds | Citation persistence at 30 and 60 days |
| Impressions-to-CTR ratio | No click data in the AI reports | AI impressions and AI referral sessions, separately |
| Blended social citations | ChatGPT 0.46% vs AI Overviews 12.2% | Social citation share per engine |
Then add the numbers leadership actually cares about: AI referral sessions and conversions from those sessions. These are the ones that survive a budget conversation. For context, Codeword makes the case that AI visibility deserves a distinct budget line in 2026, which only works if the metrics you attach to it hold up under scrutiny.
Where the real numbers come from
You don't need a dozen tools. You need first-party analytics for the traffic side and one AI visibility platform for the prompt side.
For traffic, GA4 can segment sessions from AI referrers.
For the prompt side, Promptwatch is the one I would put first for teams that want more than a tally. It tracks real ChatGPT, AI Overviews, Gemini, and Perplexity interfaces, separates the metrics above (visibility, share of voice, position, sentiment, citation rank, self-citation rate), and adds IP-verified AI crawler logs and visitor analytics, so you can tie citations to actual sessions instead of stopping at mentions.
Other tools in the category worth a look include Profound, Peec AI, and Otterly.AI, which are mainly monitoring tools. The GEO software directory at bestgeosoftware.com lists more if you want to compare.
Profound
Otterly.AI
A quick way to choose: if the tool can only tell you whether you were mentioned, it will keep generating the metrics above. If it can show crawler activity, citations by page, and traffic, you can report things that change decisions.
How to retire a metric without a fight
Telling a CMO that last quarter's number was misleading is awkward. A few things that help.
- Keep the old metric for one more cycle, labeled as "diagnostic only," next to its replacement.
- Explain it with one fact. "ChatGPT cites about five sources per answer and AI Overviews about ten" ends the blended-count conversation in a sentence.
- Show what decision each new metric drives. Persistence by content type, for example, tells you which pages need refreshing.
- Put definitions in the footer. Every number should say what it counts, which engine, and which dates.
And date everything. These engines change monthly. A chart without a period is an invitation to misread a platform change as a performance change.
A note on checking your own claims
Two caveats from my own research. First, several of the numbers above come from a single week or a single month, so treat them as a snapshot of that period. Second, I wasn't able to verify the often-quoted Gartner forecast that search volume will fall 25% by 2026. It dates from 2024, and I would leave it out of any deck until you have fresh data.
The short version
The metrics worth retiring share a pattern: they are easy to produce, they go up, and they don't tell you what to do. The ones worth keeping are slightly harder to collect, and each points at an action: fix a crawl error, refresh a decaying page, close a coverage gap, or defend a conversion path. Cut the first group from the 2026 deck, and the second group will earn the budget conversation by itself.