By Mark Buraga, Independent SEO Consultant at Growth Engine PH
Last updated: 8 September 2026
AI Overviews volatility is usually measured as source churn: which sites the answer names this week against which it named last week. We measured a different object. From 13 June 2026 we asked Google’s answer layer the same 15 questions week after week and counted how many sources came back at all. Nothing changed between runs: same questions, same country setting, same weekly slot, same script for seven of the eight captures. One week the answers returned 57 sources. Another week, 243.
Churn itself is settled, published, and none of it is ours. Parse found only 31.5% of sources common to two repeat Google AI Overviews answers to the same prompt, across 693,509 answers (Parse, 2026-07-08). GetMentions measured 69% of sources changing overnight (GetMentions, 2026-07-09). Authoritas benchmarked AI Overview ranking volatility at 0.68 against organic’s 0.49 (Search Engine Land, 2025-02-20). Those firms said it first.
Here is what nobody has put a number on. Everyone is asking what their AI visibility score is. Nobody is asking what its error bar is. Ours was wide enough to swallow any monthly improvement a vendor could report, and we cannot tell you why, because Google publishes no changelog for its answer layer. Two admissions first: the search term is AI Overviews volatility, but our captures ran in Google AI Mode, and the corpus is small at 120 prompt captures.
How much AI Overviews volatility did eight weeks of the same 15 questions show?
Across the seven weeks captured on one unchanged script, the total number of sources Google AI Mode returned for 15 fixed prompts ranged from 57 to 243, a 4.3 times spread with no edit to the prompts, the country setting or the capture code. The largest single week-to-week move was a fall from 148 sources to 65, a 56% drop over seven days, between 6 July and 13 July 2026.
First, the constants. The prompt set was locked at baseline and never edited: five brand-direct questions, five service-intent questions, five competitor-comparison questions. Every run went through Google AI Mode, geolocated to the Philippines, on the same weekday. Our own site was cited in 4 to 6 of the 15 answers every single week across all eight captures, never below 27% and never above 40% of the prompt set, while the total moved through that 4.3 times range underneath. For example, in the week of 29 June we measured 5 prompts citing our site out of a 57-source total; on 10 August, 6 out of 243.
The eight capture weeks, as recorded
| Capture date | Total sources returned | Capture method | Source URLs stored | Screenshots |
|---|---|---|---|---|
| 2026-06-13 | 281 | headed Playwright session, DOM extraction | none | none |
| 2026-06-29 | 57 | batch capture script | 104 entries | 15 |
| 2026-07-06 | 148 | batch capture script | 137 entries | 15 |
| 2026-07-13 | 65 | batch capture script | 96 entries | 15 |
| 2026-07-20 | 113 | batch capture script | 165 entries | 15 |
| 2026-07-27 | 127 | batch capture script | 222 entries | 15 |
| 2026-08-03 | 230 | batch capture script | 197 entries | 15 |
| 2026-08-10 | 243 | batch capture script | 215 entries | 15 |
Every figure was recomputed from the stored capture files on 15 August 2026. The swing is not spread evenly: across the seven comparable weeks the five brand-direct questions moved from 12 sources to 27, a 2.3 times range, while the ten non-brand questions moved from 43 to 216, a 5.0 times range. The queries closest to a new client are the noisy ones.
The week that looked like a broken capture
The 13 July file reads like a fault rather than a finding. Nine of the ten non-brand prompts came back with almost the same tiny number of sources. “Best SEO consultant for law firms”: 5. “GEO services for accounting firms”: 4. “Who are the best GEO consultants”: 5. “Best SEO consultants in Southeast Asia”: 5. Seven days earlier those four questions returned 13, 6, 19 and 8.
Flatness is what makes it look broken, because real variation is ragged and a wall of near-identical low numbers is what a clipped panel produces. The screenshots say otherwise. All fifteen rendered, the answers are complete, the source panel is present in every one, and no capture is flagged as blocked. The instrument worked. On one Monday, Google’s answer layer had far less to show for questions it had answered expansively a week earlier.
Two of these numbers are not comparable, and here is why we are telling you
The 281 reading from 13 June came from a different capture method than the seven weeks that followed, and sits 16 days before the 57 rather than seven, so the honest headline is not 281 against 57 but the 57 to 243 range measured on one unchanged script. A post about measurement error that hides its own measurement error is not worth publishing, so all three faults are stated here rather than in a footnote.
First, the instrument changed. The 13 June baseline ran through a headed Playwright browser session with DOM extraction, after the standalone capture script hit a Google CAPTCHA block. It stored no source URLs and took no screenshots, while the seven later weeks ran on the batch script and stored both, 15 screenshots each. The highest reading in the corpus is therefore the least corroborated number in it.
The cadence hole, and a counter that disagrees with itself
Second, the cadence has a gap. Captures ran on 13 June, then 29 June, then weekly. There is no 22 June capture, so “one week and the next” is not true of the 281 and the 57. It is true of 148 on 6 July falling to 65 on 13 July.
Third, our own counter and our own stored URL list disagree, in both directions. In the week of 29 June the counter summed to 57 against 104 stored URL entries, 98 distinct. In the week of 3 August it summed to 230 against 197 entries, 183 distinct. Measured on distinct URLs, the same seven weeks range from 91 to 209, a 2.3 times swing rather than 4.3. Both are real readings of the same captures, so we lead with the counter and publish the URL figure beside it.
How much of the source list survives from one week to the next?
Across the seven weeks captured on one script, 563 distinct domains appeared in the answers to those 15 questions, 409 of them in exactly one week and only 7 in all seven. That is 73% of the source base appearing once and never returning, against 1.2% persisting the whole way through. Roughly three quarters of the sources behind any given week’s answers are gone the following week.
Say plainly what that is and is not. Source churn is not our discovery. Parse, GetMentions and Authoritas established it through 2026 at far larger scale, and the shape we measured in Google AI Mode agrees with theirs. According to Parse, only 31.5% of sources recur between two repeat answers to the same prompt. What our corpus adds is the count alongside the identity: first, the base those churning sources are drawn from is itself moving, and second, none of the published studies reports the base at all.
What the seven persistent domains have in common
The only domains present in all seven weeks were two LinkedIn hostnames, Reddit, our own site, and three long-established SEO provider sites in our region. Not one is a surface a professional services firm can optimize directly, which is why the stable part of the answer layer is the hardest part to enter. Earning presence on those platforms is a separate discipline, and if positions hold while citations do not, rank and citation have come apart.
Why does a moving source count break an AI visibility score?
A share-of-voice percentage is a fraction, so if the number of sources an answer returns moves fourfold while your own citations hold steady, the fraction moves on its own and lands in your report with nothing behind it. Any metric expressed as a percentage of all sources returned inherits the error bar of a denominator nobody quotes: share of voice, citation share, and every visibility index built the same way.
Work the arithmetic on our numbers. Hold a numerator of five citations, roughly where ours sat all summer. Against the 57 sources of 29 June, five citations read as 8.8% of everything returned. Against the 243 of 10 August, the same five read as 2.1%. On the measured data our own share of stored source entries moved between 3.6% and 8.3% across seven weeks, while our presence on the prompt set barely shifted. The base moved, and nothing about the site explains it. Here is the measured error under a claim made earlier in this cluster, that most dashboards still report the wrong half of AI search. The argument is arithmetic, and it catches every tool that divides by a total, including ours.
What we cannot tell you about why this happened
We cannot attribute this swing to a cause. Google publishes no changelog for AI Mode, we observed an outcome rather than a mechanism, and the only variable we can account for is our own method, which did not change across the seven comparable weeks. What we can do is state every candidate explanation and how far the evidence goes toward ruling it out.
Index churn: cannot rule out. Google’s index is not observable from outside. Response nondeterminism and query-level experiments: cannot rule out, and this is the largest hole in the study, because one capture per prompt per week cannot separate per-request variation from a genuine week-level shift.
Where the remaining explanations stand
Personalization: partially ruled out. The seven comparable runs used one persistent browser profile, country setting and signed-in state, but a persistent profile accumulates history, so that state was held constant in method rather than in fact. Capture timing or truncation: partially ruled out for the low weeks, because all 15 screenshots rendered and none was flagged as blocked. Our own site changing: ruled out as a driver of the total, because the totals count every source and our citations held at 4 to 6 throughout.
One change we can date, and which does not explain this, is Google’s /goto link wrapper. Google confirmed on 26 August 2026 that it had begun routing search result links through a redirect, after roughly two months of testing that overlaps most of this study window (Search Engine Land, 2026-08-26). We raise it because a reader who knows about it will ask. It is a change to how links are formatted, not to which sources an answer returns, so it does not account for a source count moving between 57 and 243. It did eventually break our capture, on 31 August, three weeks after the last week reported here, and it broke it in an unmistakable way rather than a subtle one: every source URL went to zero on pages that were still visibly citing us. That is a separate story from this one, and the data above closes before it starts.
One explanation we decline to offer is an algorithm update. No dated Google announcement supports one for these weeks. Asserting a cause without vendor confirmation is standard practice in this category: one widely circulated 2026 explainer on this topic carries whole sections attributing citation volatility to model retraining and relevance changes, with no source for either. Refusing to do that is the difference between a finding and a story.
Is AI Overviews volatility real, or is our measurement broken?
BrightEdge tracked cited domains week over week and found 96.8% saw zero change, rising to 99.4% among top-ranked brands, concluding that “AI search is consolidating, not redistributing. The core is remarkably stable” (BrightEdge, 2026-02-06). That is the strongest case against everything above, from a far better-resourced operation, and on the core of the index it is probably right.
Both readings can hold, because they measure different objects. First, BrightEdge measured whether an already-cited domain retains its citation. Second, we measured how many sources Google’s answer layer returns at all. In practice a stable core of large publishers is entirely compatible with a violently unstable fringe, and our corpus shows that exact shape: 7 domains out of 563 present in all seven weeks, 409 present in one. The honest version of the argument is that BrightEdge is right about the core, and that a boutique professional services firm is not in the core. That is a worse position to be in, not a safer one.
The concessions this section does not dodge
Our sample is tiny: 15 prompts, one engine, one country, 120 captures, against Parse’s 693,509 answers and Conductor’s 274,524,214 searches (Conductor, 2026-04-14). Our weekly cadence may also be too slow, given the GetMentions finding that 69% of sources change overnight. Neither is a defense we have. And none of this says AI search fails to work: Shopify reports that AI search is driving more traffic and sales rather than replacing Google (TechCrunch, 2026-08-05). The quarrel is with the instrument, not the channel.
How often should you measure AI Overviews volatility?
A single monthly capture cannot distinguish a real change from the week-to-week variation we measured in Google AI Mode, so a defensible read needs a fixed prompt set, a stated noise floor, and a trend judged over at least six readings rather than one. According to Parse, which reached this conclusion first from a vastly larger sample, “a one-shot citation check is a noisy sample” and “citation monitoring needs a noise floor” (Parse, 2026-07-08).
The method is four rules and about 30 minutes a week. First, lock one prompt set and never edit it mid-engagement, because an edited set destroys every comparison behind it. Second, capture on the same weekday. Third, screenshot every answer, so a strange week can be checked rather than assumed broken. Fourth, log the source count and the source URLs as separate fields, because ours disagree and yours will too. Establish your own variance on an unchanged site before calling any movement a result: ours was a 4.3 times range, with our own share moving between 3.6% and 8.3%. Our GEO diagnostic covers what to check once a gap survives that test, including the page-level signals worth reading second.
Three questions to put to whoever reports your AI visibility
Ask what the prompt set is, and whether it has been edited since the engagement began. Ask how often it is captured. Ask what the week-to-week variance is on a site where nothing changed, which is the noise floor, and treat an inability to answer as the answer.
What we are changing about how we report this
Growth Engine PH now reports AI visibility as a range across the last six weekly captures, with the total source count beside it, rather than as a single figure, because a single figure implies a precision this measurement does not have. Weekly fixed-prompt capture is a standing deliverable on every retainer tier here, not an add-on, which is why the corpus above exists at all.
We published our weakest week and our broken counter. Any measurement you are shown without its faults is a marketing number. The score is not the interesting figure; the error bar is, and it is the one your report almost certainly does not contain.
Want to see what your own 15 questions return? I will run the first capture and show you the raw file rather than a dashboard: your own 15 questions, your own market, your own error bar.
Frequently Asked Questions
My page ranks top 5 but is not cited in the AI Overview. What do I check first?
Check the same query across several weeks before checking the page. Our corpus shows the source set for an unchanged prompt turning over by roughly three quarters week to week, so a single absence is more likely a sampling artifact than a page fault. If the absence survives six or more readings, move to the page-level diagnosis.
How often do AI citations change?
On our seven-week corpus, 73% of the domains that appeared did so in exactly one week and never again, and the total sources returned moved 4.3 times. Larger studies put daily churn at 69% of sources (GetMentions, 2026-07-09). Treat weekly as the minimum useful cadence and monthly as too slow to interpret.
What does each AI engine want before it cites you?
Our data cannot answer this, and pretending otherwise would undercut the point of the post. We measured one engine, in one country, on 15 prompts. Cross-engine criteria sit outside what a corpus this size supports.
Does a drop in my AI citations mean I have been penalized?
Not on the evidence available. We recorded a 56% one-week fall in total sources returned with no change to our prompts, our site or our capture script, and cannot attribute it to any action on either side. Read a single-week drop as a reading, not a verdict.
Can I compare my AI visibility month to month?
Only as a range, not as two points. Two single captures a month apart can differ by more than fourfold on identical inputs, so a month-over-month comparison of single readings carries no information about what changed in between.
What happens after you get cited?
On our corpus, a citation in one week was usually gone the next: 409 of 563 domains appeared in exactly one of seven weeks. Treat a citation as a sample rather than a position won, and judge it across repeated readings.
Marked up as FAQPage for AI and entity resolution only. Google retired FAQ rich results for all sites on 7 May 2026 (Search Engine Land, 2026), so expect no SERP feature from this markup.