Key takeaways
- No AI Mode rank tracker reads a fixed position. Google AI Mode has no public API for rank tracking, so tools drive real browser sessions and report a modeled estimate from repeated probes.
- Repeat-run variance is large. In EWR Digital's 2026 study, a single run carried a 95% confidence interval of about plus or minus 33 percentage points. Seven to eight runs brought it down to about plus or minus 10.
- SISTRIX data, reported secondhand by Kevin Indig, says AI Mode replaces 56% of its cited sources every week. Monthly check-ins are too slow.
- I could not find a published variance test that isolates AI Mode, or a head-to-head test of AI Mode trackers. Treat any vendor claiming a precise AI Mode "rank" with suspicion.
- The practical fix: more prompts, 5+ repeats each, platforms tracked separately, trends over snapshots, and confidence intervals shown next to every number.
Why this question matters now
AI Mode is no longer a side experiment. According to a Digital Applied recap of Google I/O on May 19, 2026, AI Mode reached 1 billion monthly users and AI Overviews 2.5 billion. That recap is secondhand, not a Google release, but the direction is clear. People are asking questions in AI Mode, and marketers want to know whether their brand shows up.
So they buy a tracker. The dashboard shows "visibility: 34%" or "cited in 6 of 20 prompts." Next week it says 21%. Did something break, did a competitor surge, or did you just ask twice?
That last possibility is the uncomfortable one, and the data says it is more likely than most dashboards admit.
What an AI Mode rank tracker actually measures
Traditional rank tracking is simple. A keyword returns ten blue links, and you sit at position 4 or you do not. AI Mode does not work like that. It reads a question, fans it out into sub-questions, and writes one synthesized answer with a handful of inline citations.
So a tracker is really measuring things like:
- Whether AI Mode appears for the prompt at all
- Whether your brand is named in the answer
- Whether your domain is cited as a source
- Who gets cited instead of you
Named and cited are different. AI Mode can describe your product without linking to you, and only the link drives traffic. In EWR Digital's data across five other AI engines, brands were named in 25.7% of answers and cited in 18.0%. Same brands, same prompts, a gap of almost eight points.
Then there is the plumbing. GEO Toolbox points out that Google AI Mode has no public API for rank tracking. Every tool samples it by driving a real browser session and reading what comes back. Each number is a modeled estimate from repeated probes, not a meter reading of what real users saw.
That is not a flaw in any one vendor. It is the nature of the surface. But it changes how you should read the output.
What repeat-run variance looks like in practice
Here is the awkward part of my research. I found no published repeat-run study that isolates Google AI Mode alone. What exists are studies of neighboring engines and one drift study that includes AI Mode. Together they paint a consistent picture.
The EWR Digital study
EWR Digital published a study on August 29, 2026 (with a public correction on September 5, 2026). It covered 12,778 answers across ChatGPT, Perplexity, Gemini, Claude and DeepSeek, with 118,050 citations collected over 164 days between January 13 and June 25, 2026, across 7 commercial domains. It does not include AI Mode or AI Overviews.
The method used Jaccard overlap on the set of cited domains. Median repeat depth was 15 to 24 observations per prompt and engine. When you ask the same prompt again, here is how much the cited domains overlap:
| Engine | Median overlap of cited domains | Identical citation set |
|---|---|---|
| Claude | 0.417 | 1.70% |
| Perplexity | 0.333 | 0.59% |
| DeepSeek | 0.333 | 2.48% |
| Gemini | 0.231 | 0.49% |
| ChatGPT | 0.133 | 0.06% |
Read that last row again. ChatGPT returned an identical citation set on a repeat ask in 0.06% of cases. Even the most stable engine, Claude, did so under 2% of the time.
AI Mode is not in this table, and I would not assume it sits better or worse than Gemini without data. Google's synthesis is also fan-out based, so I would not bet on it being calm.
Month-to-month movement is mostly noise
This is the finding that should bother anyone who sends monthly visibility reports. EWR found consecutive months overlapped at 0.318 to 0.458, while repeated asks within the same period overlapped at 0.309. Their conclusion: a month of apparent movement is statistically indistinguishable from asking twice.
I would soften that a little. It is one study, on seven commercial domains, on five engines. But it matches what practitioners see. A line chart that wiggles up 12 points is often just sampling.
How many runs you need
EWR's confidence intervals are the most useful practical output:
| Runs per prompt | Approx. 95% CI |
|---|---|
| 1 | about plus or minus 33 points |
| 3 | about plus or minus 18 points (standard error 0.093) |
| 7 to 8 | about plus or minus 10 points |
Their recommendation is at least 7 to 8 runs per prompt, with the interval reported next to the number. One run tells you almost nothing. A tracker that runs each prompt once per day gives you a daily number with an error bar wider than the signal you care about.
A correction worth noting: EWR originally claimed that named brands were 2.7x more stable than cited brands. They withdrew it. Run-to-run agreement was about the same for both (0.867 vs 0.879). So do not assume mentions are a more reliable metric than citations just because they feel sturdier.
Sampling variance inside a single model
Kevin Indig's Growth Memo, published via Search Engine Land in June 2026, collects more numbers. Within-LLM sampling variance reaches 10 to 34% on identical prompts (arXiv 2601.21339). In a study of 815,000 prompt-page pairs with AirOps, only 2.2% of citations remained after running the same prompt three times in ChatGPT. The same article also cites 2.3% from "The Consensus Gap," so the figure wobbles slightly depending on the source. Either way, it is tiny.
He also reports that high versus low reasoning opens an 18-point citation-rate gap and fires 4.6x more fan-out queries. Which means two trackers can both be "correct" and disagree, because they triggered different reasoning depths.
I have not read the arXiv paper or the AirOps study directly. These are secondhand figures, and I would cite them as such.
Citation drift: the AI Mode specific data point
The closest thing to an AI Mode stability number comes from SISTRIX. They tracked 82,619 prompts over 17 weeks. According to Indig's summary, AI Mode replaces 56% of its cited sources every week, compared with 74% for ChatGPT. I did not scrape the SISTRIX post, so treat the figures as secondhand and check the original at sistrix.com before quoting them.
If accurate, that means more than half of AI Mode's sources turn over in a week. A monthly tracker is sampling a different cast of characters each time. Indig's conclusion is that monthly tracking is too slow.
Why the baseline moves even when the tool is right
Variance is not the only problem. The ground itself shifts. Promptwatch Data has a good illustration in its Google AI Mode Citation Share report for June 2026, covering May 31 to June 26, month to date:
- google.com made up 7.31% of all AI Mode citations, more than YouTube (2.88%) and Reddit (2.52%) combined.
- google.com rose from 4.24% in May to 7.31% in June, a 72% increase on top of a sixfold jump the month before.
- YouTube fell from 3.42% to 2.88%, its first decline in that series.
- Below the top three, no independent website clears 1%.
So even a perfectly calibrated tracker would see your citation rate drift as Google cites itself more. Some of what looks like "we lost visibility" is the pie being re-cut. That is a market-level effect, not a sampling effect, and you need to separate the two.
Another Promptwatch report, Average Sources Per Response (published September 2025), shows how differently engines behave. Google AI Overviews cites about 10 sources and holds steady, Perplexity cites almost exactly 10 every day, and Copilot swung from under 2 to nearly 17 within a few weeks. The report has no AI Mode figure and no repeat-run variance data, so I would not stretch it. But it makes the point: stability is a property of the engine, and you cannot assume it.
There is also a calendar problem. Google announced a Core Update on May 21, 2026, two days after I/O. Any variance test run around that date has extra volatility baked in.
Other sources of error in AI Mode tracking
LLM Pulse lists the main accuracy caveats for AI Overview checkers, and they apply equally to AI Mode:
- Personalization: a clean session may differ from what your real users see.
- Live rewrites by Google.
- Regional rollout gaps.
Their advice is to run checks from multiple geos over multiple weeks, especially for high-stakes decisions.
And a vendor-side caution from GEO Toolbox: when a vendor will not name the surface it tracks, assume it does not track it. "AI search" in a feature list might mean ChatGPT only. Ask specifically whether AI Mode is covered, and how.
How to run a tracking setup that you can trust
The common setup Indig criticizes is 25 to 50 prompts, run once per platform, tracked daily. Based on the numbers above, that setup produces a lot of data and not much information.
His worked example is a better template:
- 40 seed prompts: 12 brand, 12 category and 16 problem prompts.
- 5 repetitions per prompt per platform, run weekly.
- Platforms tracked separately, never blended into one "AI visibility score."
- Prompts customized for 3 personas.
- Multi-turn follow-ups tracked too.
I would add a few of my own rules:
- Report an interval, not a point. "Cited in 30% of runs, plus or minus 10" is honest. "30%" alone is not.
- Compare periods only when the change exceeds the interval. If your CI is 10 points and you moved 6, you moved nothing.
- Keep location and language fixed inside a monitor so markets do not mix.
- Pair tracker data with first-party data (below).
- Log the date of known Google updates and annotate your charts.
Use first-party data as a check
Google Search Console began providing AI Mode impression data in June 2025, according to a community thread on r/GEO__AI__SEO. Google also announced gen-AI performance reports in June 2026 (developers.google.com/search/blog/2026/06/gen-ai-performance-reports), and Bing Webmaster Tools launched an AI Performance report in public preview in February 2026. Neither gives you competitor citations. But they are real user data, which no sampled tracker can claim to be.
If your tracker says visibility doubled and Search Console impressions are flat, believe Search Console.
Choosing a tracker: what to ask vendors
I did not test trackers head to head, and I have no current pricing. So this is a question list, not a ranking. Ask any vendor:
| Question | Why it matters |
|---|---|
| Do you track AI Mode specifically, and from the real interface? | API output and user-facing answers can differ. |
| How many runs per prompt per reporting period? | One run has a roughly plus or minus 33 point interval. |
| Do you show confidence intervals or variance? | Without them you cannot tell signal from noise. |
| Can I choose country, language and location? | Regional gaps change what AI Mode shows. |
| Can I run more often than daily? | Faster cadence helps follow launches and fast drift. |
| Do you separate named from cited? | They are different outcomes with different traffic impact. |
| Can I export raw responses? | You should be able to audit the samples. |
Some tools in this space to look at, depending on your situation:
Promptwatch monitors the real interfaces of Google AI Mode, AI Overviews, ChatGPT and others by country and language, and a monitor can run every 12, 6 or 1 hours on Business and above, which gives you more samples per period. It also adds crawler logs and visitor analytics so you can check tracker numbers against what actually hit your site.
Other options that are in our catalog and claim AI Mode or AI Overviews coverage include SEOTalos, Otterly.AI, and Peec AI. I have not verified their repeat-run methods, so ask the questions above before buying.
Otterly.AI
For a wider view, the directories at ai-rank-tools.com and bestgeosoftware.com list many more trackers.
What to tell your boss or client
The honest line is this: AI Mode visibility is a probability, not a position. You can estimate it, and the estimate improves with more runs. But a single check, or even a week of single checks, will mislead you.
Practical reporting rules:
- Lead with trends over four or more weeks, not week-over-week deltas.
- Show the interval.
- Treat a competitor suddenly "overtaking" you with suspicion until it shows up across several repeat runs.
- Judge success by the outcomes you can verify: traffic, conversions and Search Console impressions.
Trackers are still worth having. The sources that get cited repeatedly across many samples are the ones worth studying, and the pages that never appear are the ones worth fixing. You just have to stop reading the dashboard like a thermometer.
Gaps in the evidence
To be clear about what I could not establish:
- No published repeat-run variance study isolates Google AI Mode alone. SISTRIX is the closest, and I only have it secondhand.
- I found no head-to-head test of AI Mode trackers.
- AI Rank Lab ran 968 repeated prompt runs, but I only saw a search snippet and have not verified the content.
- Promptwatch has no report that measures repeat-run variance directly.
If you run your own test, five repeats per prompt over four weeks on 40 prompts is enough to see where your noise floor is. That number is worth more than any vendor's accuracy claim.