What we did
We took the 200 commercial Perth searches from the Perth AI Search Study 2026, covering 20 industries, each run three times on three separate days between 8 and 10 September 2026. That is 600 Google searches. Google returned an AI Overview on 393 of them and cited 2,845 sources in total.
For every one of those citations we recorded the URL, the page it pointed at, and what kind of source it was. Then we compared each citation to Google’s ordinary organic results, captured in the same request that returned the AI Overview, so the search, the day, the location and the device match exactly rather than approximately.
No new data was collected for this report. Everything here is a second reading of the parent study’s records, and it cost nothing to produce.
This report is published by AI SEO Perth, a research publication operated by Search Scope. Search Scope is an SEO agency in Perth, and “SEO agency” is one of the 20 industries measured. Search Scope’s own pages were cited 6 times out of 2,845, on 2 of the 200 searches, which places it nowhere near the most-cited domains below. We are stating that at the top rather than the bottom.
This page exists as evidence, not because anyone searches for it. Our own Search Console records 1,795 impressions for “ai seo perth” and 1,366 for “ai search optimisation perth” since 1 June 2026, at average positions of 23.9 and 49.5. There is no measurable demand for a question about AI Overview citation sources specifically, and we are not claiming any.
Most of what it cites was already ranking
Set aside the citations that point at Google itself, and 57.6% of what a Perth AI Overview cited was the exact URL of a page already ranking in Google’s own organic results on the same page for the same search.
| Where the cited page sat in the organic results on the same page | Citations | Share |
|---|---|---|
| Exact URL, organic positions 1 to 3 | 600 | 25.0% |
| Exact URL, organic positions 4 to 10 | 783 | 32.6% |
| Same domain ranking, but a different page cited | 181 | 7.5% |
| Not observed in the organic results we captured | 839 | 34.9% |
That is 2,403 citations to ordinary web pages. Counting all 2,845 citations, including the ones pointing at Google’s own properties, the exact-URL figure is 48.6%.
The detail that surprised us is the second row from the bottom. When the AI Overview cited a domain that also ranked on page one, it cited the exact ranking URL 88.4% of the time, 1,383 times out of 1,564. Citing a different page on a ranking site is the rare case, not the common one.
Read the same data from the other direction and the gradient is shallower than you would expect. Of every organic result Google captured at position 1, the AI Overview on that same page cited it 64.1% of the time. By position 8 that had fallen to 32.3%.
| Organic position | Cited by the AI Overview on the same page |
|---|---|
| 1 | 64.1% |
| 2 | 51.7% |
| 3 | 39.3% |
| 4 | 44.7% |
| 5 | 40.8% |
| 6 | 34.3% |
| 7 | 39.2% |
| 8 | 32.3% |
| 9 | 31.7% |
Almost all of the drop happens between positions 1 and 3. From position 4 down to position 9 the citation rate sits in a narrow band between 32% and 45%. Being eighth is worth roughly half of being first here, not a tenth.
One caveat travels with every number above, and it is load bearing. We captured Google’s first page of organic results only, a median of eight results per search. A cited page that did not match is recorded as not observed in those results. It is not evidence that the page does not rank at all, so 57.6% is a floor and the true overlap is higher by an amount we cannot measure.
Nothing here is causal. Ranking and being cited are two things the same system does on the same search at the same moment. We have no experiment that separates them, and neither does anyone else reading this data.
But it does not always appear, and what it cites keeps changing
Before any of that is useful, two instabilities have to be on the table.
An AI Overview did not reliably appear. Of the 200 searches, 91 produced one in all three rounds, 71 produced one in some rounds and not others, and 38 never produced one at all. Roughly a third of these searches were unstable at the level of whether an AI answer appeared, before anyone looks at what was inside it.
When one did appear, its bibliography moved. We compared the cited sources for every pair of rounds where an AI Overview appeared in both, 322 pairs across 140 searches. We never treat a missing AI Overview as an empty citation list, because the two are different phenomena.
| At URL level | At domain level | |
|---|---|---|
| Average overlap between two rounds | 0.43 | 0.48 |
| Pairs sharing at least one source | 94.1% | 94.1% |
| Pairs sharing nothing at all | 5.9% | 5.9% |
Those overlap figures are Jaccard scores, where 1 means the two rounds cited exactly the same set and 0 means they shared nothing.
Among the 91 searches that produced an AI Overview on all three days, only 23.5% of cited URLs appeared in all three, and 55.9% appeared in exactly one round. At domain level it is steadier: 32.9% appeared in all three.
The practical reading is that the AI Overview stays recognisably the same answer while swapping out the pages it credits for it. Checking once whether Google’s AI Overview cites your page is taking a single reading of a number that moves.
Nearly half of what a business gets is its homepage
Of the 1,861 citations that went to a business’s own website, 891 of them, or 47.9%, were the bare homepage.
| What was cited on the business’s own site | Citations | Share | Already ranking as the exact URL |
|---|---|---|---|
| Homepage | 891 | 47.9% | 67.7% |
| Service or product page | 555 | 29.8% | 44.7% |
| Article or guide | 142 | 7.6% | 38.7% |
| Location or suburb page | 130 | 7.0% | 52.3% |
| The business’s own “best of” style page | 129 | 6.9% | 51.2% |
| Everything else | 14 | 0.8% | mixed |
Looked at per business rather than per citation, 499 business domains were cited at least once. Of those, 164 were only ever cited at their homepage, 252 were never cited at their homepage, and 83 were cited both ways.
Homepages were also far more likely to be a ranking page in their own right: 67.7% of homepage citations were an exact organic match, against 45.8% for every deeper page combined.
Two readings fit that and we cannot choose between them with this data. The homepage may simply be the strongest ranking URL a small business has for a head commercial search. Or the system may prefer an entity-level page when the question is “who should I use”. Both are guesses, and we are labelling them as guesses.
What it actually cites
| Source type | Citations | Share | Domains | Searches it appeared on |
|---|---|---|---|---|
| Business’s own website | 1,861 | 65.4% | 499 | 160 |
| Google’s own properties | 442 | 15.5% | 2 | 32 |
| Forums and user discussion | 151 | 5.3% | 2 | 73 |
| Comparison and “best of” pages | 122 | 4.3% | 21 | 53 |
| Directories and listing sites | 114 | 4.0% | 12 | 45 |
| Social networks | 61 | 2.1% | 7 | 31 |
| Government and regulators | 37 | 1.3% | 3 | 17 |
| Unclassified | 24 | 0.8% | 15 | 17 |
| Review platforms | 10 | 0.4% | 3 | 8 |
| Industry associations | 9 | 0.3% | 5 | 6 |
| Software and tools | 9 | 0.3% | 1 | 4 |
| News and editorial media | 5 | 0.2% | 1 | 2 |
The parent study reported this same split two ways, strictly at 38.2% and broadly at 70.9%, the difference being whether a site had to be matched to a business we had already identified or merely had to look like a business website. The 65.4% above is the broad figure after we corrected three classification faults in our own work, which we describe at the end of this report.
Two rows are worth pausing on.
News and editorial media are effectively absent. Five citations out of 2,845, one URL, two searches. Whatever a Perth AI Overview is built from for a commercial question, it is almost never a news story. We are reporting that as a null result because null results are what stop a study being a sales argument.
Google cited itself 442 times. 427 of those were Google Business Profile panels, the business cards Google hosts rather than any page on the open web. They appeared on 32 of the 200 searches, and 9.6% of citing AI Overviews contained at least one. These citations cannot appear in organic results by construction, which is why every overlap figure in this report is published twice, once over everything and once over ordinary web pages alone.
Reddit reaches more Perth searches than any other source
| Domain | Searches it was cited on | Citations | Industries | What it is |
|---|---|---|---|---|
| reddit.com | 73 | 147 | 19 | Forum |
| google.com | 26 | 436 | 13 | Google’s own panels |
| facebook.com | 25 | 48 | 14 | Social |
| canstar.com.au | 10 | 13 | 3 | Comparison site |
| westernpower.com.au | 9 | 19 | 1 | Government-owned utility |
| studiohawk.com.au | 9 | 18 | 1 | Business website |
| perthsolarwarehouse.com.au | 8 | 19 | 2 | Business website |
| pwd.com.au | 8 | 17 | 1 | Business website |
| seotakeoff.com | 8 | 15 | 1 | Comparison article |
| regenpower.com | 8 | 14 | 1 | Business website |
We rank this table by how many different searches a domain reached, not by how many citations it collected. A domain cited three times on one search because we asked that search three times has not reached three searches, and ranking by raw count rewards exactly that. The automatically generated table further down this page counts something different again, the number of individual observations citing a domain, so its figures are higher than the search counts here by design.
Reddit reached 73 of the 200 searches across 19 of the 20 industries. No other non-Google domain reached more than 10. Reddit and Facebook together were cited on 87 of the 200 searches. Reddit threads were also unusually well aligned with the organic results, matching exactly 70% of the time, which is consistent with those threads ranking on page one in their own right.
Underneath that, the tail is long and thin. Of the 571 domains cited, 326 were cited on exactly one search and 171 were cited exactly once in the whole study. The ten domains above account for a large share of citations but still only reach 115 of the 200 searches between them.
The same is true at page level. The single most widely cited page was one Western Power consumer-advice article, “What to look for in a reliable solar installer in Perth”, cited on nine different solar searches across all three rounds. No page in the top 20 spanned more than one industry, so breadth here is within a vertical rather than across the market.
Asking a question changed what got cited, not just whether an answer appeared
For 100 of the 200 searches we used five fixed phrasings per industry, changing only the wording while keeping the service and the location identical.
| Phrasing | Produced an AI Overview | Share of its citations that were Google Business Profile panels | Share that were the business’s homepage |
|---|---|---|---|
| “best [service] Perth” | 6 of 20 searches | 0.0% | 52.6% |
| “top [service] Perth” | 10 of 20 | 0.0% | 53.6% |
| “recommended [service] Perth” | 12 of 20 | 28.2% | 34.3% |
| “who is the best [service] in Perth” | 19 of 20 | 24.9% | 39.2% |
| “recommend a good [service] in Perth” | 20 of 20 | 21.9% | 43.0% |
The trigger-rate half of that was reported in the parent study. What is new here is the right-hand side. Google Business Profile panels were entirely absent from the two keyword phrasings and supplied between 22% and 28% of citations on the other three. Directory and business-profile pages rose from around 3% to 5% of citations to between 27% and 35%.
What did not move is how closely the citations tracked the organic results: 47.5% exact overlap for keyword phrasings against 48.1% for question phrasings. So phrasing appears to change whether an AI Overview appears and what kind of source it reaches for, while leaving its relationship to the organic results alone.
We cannot say why the panels switch on. It may be a different AI Overview template being served for recommendation-style wording. Nothing in this data identifies a template, so that stays a guess.
The differences between industries
Each industry here is only 10 searches and 30 observations, so these are indicative rather than precise, and we are deliberately not producing 20 separate findings from 20 industries.
Five differences are large enough to be worth stating:
AI Overviews appeared far more often in advice-heavy industries. Buyers agent searches returned one on 93.3% of observations and SEO agency on 86.7%. Chiropractor, electrician, plumber and veterinarian all sat at 53.3%.
Google Business Profile panels clustered in health and trades. They were 47.9% of physiotherapy citations, 39.1% of plumbing and 37.3% of reticulation, and exactly zero for SEO agency, builder, family lawyer, conveyancer, financial adviser and accountant.
The business-website share ranged from 45.3% to 88.9%. Electricians, builders, family lawyers and accountants sat at the top. Mechanics, mortgage brokers, solar installers and physiotherapists sat at the bottom, where panels and directories took the space instead.
Mechanics were the directory industry. 27.4% of mechanic citations went to directories or “best of” pages, against 8.3% across the study.
Organic overlap ranged from 30.8% to 63.0%, and the low end is explained by panel share rather than by weaker alignment, because a Google panel can never match an organic result.
What we got wrong, and fixed
Two things went wrong while producing this report, and both are worth reporting because both changed a published number.
Our first page-type classifier was wrong about Australian URLs. It treated any “-perth” or “-wa” in a slug as evidence of a location landing page. Manual review of a sample showed that reading was wrong for 19 of the 23 pages we checked, because almost every Australian service page carries a city modifier. We rewrote the rule to require a real locality signal, a suburb name or a locations section, and re-reviewed. The final classifier agrees with manual review on 90.8% of an independent 11% sample, and the remaining disagreement sits on the boundary between service pages and articles, which does not affect the homepage finding.
We found three classification faults in our own parent study. A brand-matching step was overriding the source taxonomy, so Canstar, SolarQuotes, AutoGuru, HotDoc, HealthEngine and five “best of” domains were being counted as business websites, 164 citations in all. Domain rules were matching as unanchored substrings, so driftrealestate.com.au matched realestate.com.au and schonautomotive.com.au matched automotive.com.au. Nine further domains needed a human decision, among them Western Power, which is a Western Australian Government-owned network operator publishing consumer advice rather than a business we could have recommended.
Correcting all three moves the business-website share of AI Overview citations from 70.9% to 65.4%. We have kept the original labels in the dataset alongside the corrected ones so anyone can see exactly what changed.
What we cannot tell you
We cannot tell you why any page was cited. We can tell you what was cited and what it looked like, and that the same system had those pages in its organic results at the same moment.
We did not measure schema markup, page structure, llms.txt files, word counts, freshness, brand mentions or anything about what those pages say. We have nothing to offer about them in either direction.
We cannot see past Google’s first page of organic results, so every overlap figure is a floor.
We cannot say whether any of this holds outside Perth, outside these 20 industries, outside English, or outside the three days in September 2026 when we collected it.
And we cannot tell you that appearing in an AI Overview is worth anything, because we did not measure a single click.
Provenance
Published by: AI SEO Perth
Research conducted: Search Scope
Lead researcher: Dorian Menard
Derived from: Perth AI Search Study 2026
Surface: Google AI Overviews only
Collection period: 8 to 10 September 2026
Observations: 600 searches, 393 AI Overviews, 2,845 citations
Methodology version: 2.5 parent, 1.0 derived
Additional cost: nil, no new collectionThis is a derived study. It adds no observations to the parent record and modifies none. The full collection method, including the frozen question set and its checksum, every endpoint and parameter, and the exclusions, is on the Perth AI Search Study 2026 methodology page.
The 600 observations behind this report are published alongside it, at /data/google-ai-overviews-perth.csv and as JSON. They are the same records as the parent study’s Google AI Overview rows, carried through unchanged except for the corrected source classes.
The citation-level working dataset behind the page-type and organic-overlap tables holds one row per citation, with the normalised URL, the source type under both the original and corrected taxonomy, the page type and its rule, and the organic position of that exact URL and of its domain on the same search and round.