We ran the same 100 Google searches three times. Only a quarter of the AI Overview citations showed up every time.

Seven minutes apart

At 17:09 on 21 August we ran cetaphil face wash on Google India. No AI Overview.

At 17:16, seven minutes later, same query, same market, the AI Overview was there, citing eight sources.

Nothing changed except the clock. We wondered how often this happens, so we measured it.

Both runs recorded through the same instrument, seven minutes apart. Popular Products sat at SERP position 1 on both.

Google’s answer is not stable

Across three draws, the number of queries showing an AI Overview held almost still: 63, then 63, then 60.

That looks like stability. It is not. Underneath, 56 queries showed an AI Overview every time, 32 showed one on no draw, and 12 flipped. The aggregate barely moves because the flips cancel out. Jewellery ran 9, then 12, then 9. Skincare ran 15, then 13, then 13.

Then the sharper number. Among the 56 queries that showed an AI Overview on all three draws, only 24% of the URLs ever cited appeared in all three. Median citation overlap between any two draws was 38%, ranging from complete agreement to no shared sources at all.

This is not Google being broken. It is a generative system behaving like one. The problem is that the industry reports single measurements as if there were no variance to report.

Each row is one query, each column one draw. The scattered rows are queries whose AI Overview came and went inside three days.

Brand queries are the least reliable

If you sell something and someone searches your brand name, the AI Overview is roughly three times less likely to behave consistently than on a generic query.

By category: jewellery 25%, skincare 15%, hair 10%, supplements and home & kitchen 5% each.

One case worth sitting with. dot and key sunscreen, an Indian D2C brand searching for its own name. dotandkey.com holds organic positions #1, #2 and #5. On all three draws, Google returned no AI Overview at all.

Brand-name queries are the least reproducible query type in the study.

The sites AI “trusts”, measured

There is a list going around. Reddit at number one, then YouTube, Wikipedia, LinkedIn, G2. The argument is that AI does not trust what companies say about themselves, so it reaches for forums and third-party platforms instead.

We did not set out to test that. But we had 1,391 citations sitting in a spreadsheet, so we looked.

Brand websites are the single largest source of citations in this dataset. They out-cite Reddit twenty-three to one. LinkedIn and G2 do not appear once, on any of the 100 queries, on any of the three draws.

The caveat, up front. Most of that advice is about ChatGPT. This study measured Google’s AI Overviews, in India. Those are different systems with different retrieval. So this is not a refutation, it is a measurement of a different thing. If someone tells you Reddit is the number one source for AI visibility, the honest question is: in which system, in which market, measured how, and over how many observations? We can answer all four for ours.

Every source measured across the same 100 queries and the same three draws.

Reddit is small, and unstable when it does appear

The share figure is the less interesting half. Reddit appeared on 11 of our 100 queries. But it held all three draws on only four of them. On the other seven it came and went between runs.

Where Reddit does hold steady there is a pattern: can we use body lotion on face, best deodorant for men, lab grown diamond vs natural, is giva jewellery real silver. Comparisons and is-this-true questions, where the useful answer is a human opinion rather than a specification.

Orange means cited on every draw. Grey means cited on one or two draws only.

How brands actually get cited

53% of brand-owned citations are editorial pages, blog posts, guides, advice articles. For marketplaces, only 25% are. Brands get into AI Overviews by publishing. Retailers get in by having inventory pages.

These figures cover sellers only. YouTube, Instagram and Reddit are excluded from this comparison because they have no product pages to be compared against, 281 citations or 20% of the dataset. They are included in every other analysis in this study, and YouTube is in fact the most-cited single domain in it.

The exception proves the mechanism. In supplements, brand-editorial drops to 32%, because that category’s citations are dominated by medical and institutional publishers. The more health-regulated the category, the less a brand blog can buy its way in.

The mirror image. Brands are cited for what they write, marketplaces for what they stock.

Who gets cited

Two things stand out. Brands are the largest bucket, which is more encouraging than we expected going in. And a quarter of all citations point back at Google itself, through Shopping links and YouTube.

Reddit and Quora shown separately rather than pooled as “forums”.

Indian brands cited reliably

A brand counts here only if it was cited on the same query across all three draws, on at least two such queries. Cited once, on one draw, does not qualify.

Home & kitchen has no qualifier, and the precise reason matters. Five Indian brands each held exactly one query across all three draws: the sleep company on how to clean a mattress, milton on how to clean water bottle, wakefit on which is best mattress for back pain, mafatlals on what is thread count in bedsheet, and kurlon on how to clean a mattress. None reached two. They are shown in grey below.

Sleepwell is the instructive near-miss. It was cited five times across three different mattress queries, but never on the same query twice, so it holds none. Cited often and cited reliably are not the same thing, which is the point of the whole study.

Each category contains 20 distinct queries, each run three times. Counts only queries where the brand was cited on every draw.

Citation is not a top-10 game

32% of citations came from pages ranking below organic position 10. Real web pages at positions 54, 62, 64, 67, 70 and 72. YouTube videos deeper still, out to 96.

This matters because most tools match citations against the top 10 only. Everything deeper reads as “not in top 10”, indistinguishable from not ranking at all. We matched against the full top 100.

And when the AI Overview cites a real web page, it cites the exact URL 80% of the time. Not the domain, the specific page. The unit of optimisation is the page.

Google Shopping links are excluded, since they have no organic equivalent.

What we got wrong

We recorded findings as we went. Later data killed seven of them. Publishing the corrections rather than the tidied-up version is the point, because anyone can reproduce this from the dataset below.

“Problem-intent queries always get an AI Overview.” True on draw 1, 27 out of 27, no exceptions. Across three draws, two flipped. A strong tendency, not a law.

“Thin citation sets are stable, fat ones are not.” At five observations the split was perfect. At 38 the correlation between citation-set size and overlap was +0.04. No relationship whatsoever.

“A Popular Products block at position 1 suppresses the AI Overview.” Killed by cetaphil face wash, the query at the top of this page, which had Popular Products at position 1 on both halves of its flip.

“Brands own the citations on their own branded searches.” Every confirming case shared an uncontrolled variable: the brand also held organic #1 for its own name. livon hair serum broke it, sitting at #2 and cited in none of six citations while Flipkart at #1 was cited.

The pattern in all seven: we named a mechanism from partial data and the next batch broke it. Two of our four most striking individual findings turned out to be one-draw artifacts. That is what three draws are for.

“Five citations on hair oil for hair growth were Google Shopping links.” Found on 29 August, after publication. They were not Shopping links. In late August Google began routing search results through google.com/goto intermediary URLs that redirect to the real page. Unresolved, a citation to any site records as pointing at google.com. DataForSEO had already fixed this for organic results and shipped the AI Overview fix on 28 August, four days after we published. Our collection window sat inside the gap.

We audited all 334 export files. Two were affected, six URLs in total: five citations, all on one query, and one organic result that was never cited. The five point to a PMC paper, Khadi Natural, Dr Batra’s and two Mamaearth pages, identified from the page titles in the export because the redirect payload does not decode. Five citations out of 1,391 is 0.36%. Recoding them moves one number: Google’s own surfaces fall from 26% to 25% of all citations, and the brand, editorial and stability figures are unchanged. Because the redirect payload does not decode, we cannot know where those five pages ranked, so they are excluded from the rank-match and citation-depth figures exactly as Shopping links already were. The dataset has been re-uploaded with the rows recoded, a redirect-status column added and a correction log sheet documenting every figure that moved.

Why it survived our coding is the part worth keeping. An unresolved redirect scores as “no organic match”, which is exactly what a genuine Shopping link scores. Nothing in the data distinguished them. We have since added a check that flags these explicitly instead of letting them pass as a plausible finding.

Method

How the 100 queries were chosen. We started from 24 seed keywords across five categories: sunscreen, face serum, face wash, moisturizer and face mask; protein powder, multivitamin, biotin, fish oil and creatine; shampoo, hair oil, hair serum, deodorant and body lotion; mattress, bedsheet, cookware set, water bottle and storage containers; and silver jewellery, gold plated jewellery, lab grown diamond and pearl jewellery. Each seed was expanded in Dorqa’s keyword research for India, and near-duplicate variants were collapsed into one representative query — 82 of the final 100 stand for a collapsed group rather than a single exact string. The remaining queries were sorted by monthly search volume and the top ones taken, aiming for a 500 per month floor, until each category held 20.

Two places that plan did not hold, stated plainly. Jewellery had only four seeds with enough distinct queries, so it runs four seeds of five rather than five of four. It is also the one category where the volume floor slipped: five jewellery queries fall below 500 a month, the lowest at 50. Every volume is in the published query list, so you can see exactly which. Median volume across all 100 is 13,450 a month.

Intent was labelled by hand, before collection, and the labels are ours. Each query was placed in one of three buckets: problem, commercial or brand. DataForSEO returns its own intent field and it disagrees with ours often — it calls 36 of our commercial queries and 16 of our brand queries Transactional. Both labels ship in the dataset so you can re-cut the analysis on theirs instead.

Every query run three times: 21 August (17:46 to 20:59), 22 August (13:03 to 15:25), 24 August (09:09 to 11:51). All draws forced fresh, no cached results. location_code 2356 (India), language_code en, organic depth 100. Data via DataForSEO, collected through Dorqa’s AI Overview tool.

We also checked the pilot queries by hand on Google in an incognito window. The image below is cetaphil face wash at roughly 17:20 on 21 August, four minutes after the tool recorded an AI Overview on the same query. Sponsored Products and Popular products occupy the top of the page and no AI Overview is visible in the frame. Manual checks are not the instrument, they are personalised and location-specific, but they tell us a human saw what the tool recorded.

Manual check, Google India, incognito, 21 August 2026. Bookmarks bar cropped.

Limitations

There is no way to distinguish “Google returned no AI Overview” from “the AI Overview retrieval failed”. DataForSEO confirmed this. The presence rate has a false-negative floor we cannot measure. Three draws mitigate it, and 32 queries were absent on all three.

Query selection is head-heavy. Median volume is 13,450 searches a month, so this describes high-volume queries, not all queries.

The intent mix is uneven across categories, because it fell out of what each seed produced rather than being balanced by design. Skincare and hair have five brand queries each, supplements two, jewellery one. The 29% brand flip rate rests on 17 queries in total, and the single jewellery brand query happens to be one of the five that flipped. The direction of that finding is clear; the exact percentage has a wide error bar.

Draws are unevenly spaced, T, T+19h and T+61h, across a Friday evening, Saturday afternoon and Monday morning. Deliberate breadth, but draw-to-draw figures are not directly comparable.

Draw 3 was collected under materially worse upstream latency than draws 1 and 2.

Owner bucket and page type were model-assigned. A 100-row random sample was reviewed, which surfaced three error classes: publisher articles mis-coded as product pages, platform content forced into a distinction that does not apply to it, and an incomplete Indian / international brand flag. All three were corrected and the full dataset recoded. The corrected page-type rule was then re-tested against 16 known-answer URLs and passed all 16. Brand nationality was hand-verified domain by domain. A residual error rate against the corrected coding was not counted; the sample was used diagnostically.

Full domain classification rules →

Glossary

The dataset

Everything is free, no email required.

If you disagree with a number on this page, the file is above. Check it.

Indian brands in this dataset

dot and key · giva · plum goodness · muscleblaze · foxtale · the deconstruct · iba cosmetics · pilgrim · wellbeing nutrition · tanishq · the sleep company · sleepwell · old school rituals · innovist · forest essentials · khadi natural · skinaa · mamaearth · parakkat jewels · milton · wildflower naturals · limelight diamonds · varniya · oziva · borosil · kurlon · deyga · soulflower · anveshan · ajmal · meant · citta · juicy chemistry · sceona · nuyug · wakefit · o3+ · denver for men · mcaffeine · palmonas · dr sheths · caratlane

111 Indian brand-owned domains were cited, out of 163 brand domains in total.

About the author

Hemant Sharma is the founder of Dorqa and of AdSchoolMaster, an SEO training community of 7,000+ practitioners. He has spent close to a decade in SEO and now works on AI search visibility, and earlier worked at Google and Genpact. Based in Gurugram, India, he designed, collected and coded this study himself.

LinkedIn · hemantsharma.me · Questions, corrections and disagreements are welcome.

What does this look like in your category?

We ran five. There are hundreds. If you want to see which sources Google cites in yours, and how much they move between draws, that is what Dorqa’s AI Overview tool does.

Start free →