What we did
In Google AI Mode answers, 1,932 of the 2,120 business recommendations we recorded, 91.1%, had no citation to that business’s own website among the same answer’s top-level sources. In Google AI Overviews it was 1,442 of 2,084, 69.2%. Both figures are from the registered analysis and count recommendations, not distinct businesses or answers.
Three definitions carry everything on this page:
- A recommendation is one business put forward as a provider in one answer: as a business card, in a list of providers, or named in the text as a choice. A business recommended in three answers counts as three recommendations.
- A citation is a source link at the top level of the answer that points to the business’s own website.
- Commercial questions means three of our four question types together: commercial discovery, recommendation, and comparison and trust. Informational questions are the fourth.
We sent 200 questions about 20 Perth industries to both surfaces once a day for five days, 9 to 13 November 2026: 2,000 requests and 1,942 answers. Mixing provider questions (“recommend a good SEO agency in Perth”) with information questions (“What affects the monthly cost of SEO for a small business in WA?”) let us see the same business cited on one and recommended on another.
This report is published by AI SEO Perth, a research publication operated by Search Scope. Search Scope is an SEO agency in Perth, owned by me, Dorian Menard, and SEO agency is one of the 20 industries we measured. In our September study, our pages were cited 6 times in Google AI Overviews, on 2 of 200 questions, neither of which we asked again here. In November, across 1,942 answers, our sites were cited 0 times. Search Scope appeared once, as a listing in a Google places panel in one AI Mode answer, not as a recommendation. This study began because we had seen our own pages cited without being recommended. In this data that case is not reproduced. We are stating that at the top rather than the bottom.
What the 91.1% means, and what it does not
What it means. Of the 2,120 times an AI Mode answer recommended a business, 1,932 times that answer’s top-level sources included no link to the business’s own website. The as-run analysis gives the same picture: 2,049 of 2,247, 91.2%.
It is not a share of businesses. A business recommended in ten answers counts ten times. It is also not a share of answers: one answer can recommend several businesses.
It is not the share with no reference to the business at all. AI Mode answers often list Google-hosted business references, such as a Maps listing, among their sources. We matched those references to businesses by the name the reference displays: 1,937 of the 1,999 Google-hosted references matched a business. Counting a matched reference as a citation, the share of AI Mode recommendations with no reference to that business is 34.1% (722 of 2,120).
That matching is not perfectly clean, and the inconsistency starts at the source. Each reference also links to a Google place, and in 9 of the 508 places in the data the same place appears under the names of different businesses. In one answer we checked, Google’s own response shows “Southern Cross Cleaning” linking to the place it elsewhere shows as “Nexus Kleen”. Leaving out references to those 9 places, the share with no reference is 37.6% (798 of 2,120). Also requiring the reference name to match the name the answer used for the recommendation exactly, it is 42.7% (906 of 2,120); that strictest rule also drops plain variants such as a “Pty Ltd” or “Perth” suffix. So, depending on the matching rule, between 1,026 and 1,210 of the 1,932 recommendations without an own-website citation had a Google-hosted reference to the same business in the same answer.
It does not show which source supported which recommendation. We recorded what appeared in the same answer. A Google-hosted reference and a recommendation of the same business sitting in one answer are present together; this data does not show that one rests on the other, or that the business’s website played no part.
It is slightly overstated. Our rules miss some recommendations. In a hand-checked sample, correcting them lowered the AI Mode share by about 2 points (85.7% to 83.4% in that sample). The finding holds, at a slightly lower level than shown.
How we collected it
| Questions | 200: 10 for each of 20 Perth industries |
| Question types | 60 informational, 60 recommendation, 40 commercial discovery, 40 comparison and trust |
| Rounds | 5, one a day, 9 to 13 November 2026, each posted between 16:11 and 16:15 Perth time |
| Surfaces | Google AI Overviews and Google AI Mode |
| Settings | Logged out, desktop, English, google.com.au, Perth sent as the location with every request |
| Also recorded | The organic top 10 and local pack on every AI Overview page; a Google Maps top 20 for each industry’s main search, every round |
| Collection | 2,100 requests through the DataForSEO API (2,000 AI, 100 Maps), every one returned |
| Answers | 1,942: AI Mode 1,000 of 1,000 requests; AI Overviews 942 of 1,000 |
The 20 industries were accountants, air conditioning contractors, business consultants, commercial cleaners, commercial lawyers, conveyancers, cybersecurity consultants, digital marketing agencies, electricians, home renovation builders, IT managed services, mortgage brokers, pest control, plumbers, real estate agencies, recruitment agencies, roofers, SEO agencies, solar installers and web design agencies.
An AI Overview appeared on 99.3% of informational searches (298 of 300), 94.5% of comparison and trust searches (189 of 200), 93.7% of recommendation searches (281 of 300) and 87.0% of commercial discovery searches (174 of 200). Those rates belong to this question set, not to Google search in general. Where no AI Overview appeared we recorded an absence, never an empty answer.
We wrote and registered the study protocol, ran a pilot, fixed the extraction rules and froze the question set on 3 November 2026, before any of the five rounds ran. The findings were fixed on 14 November 2026, before a word of this report was written.
How accurate the extraction is
Every answer was read by rules, not by hand, so we checked the rules against a hand-scored sample of 195 answers drawn after the rules were frozen: 645 rows in all.
| Check | Result |
|---|---|
| Rows where the rules and the hand check agree | 580 of 645, 89.9% |
| When the rules say “recommended”, it was | 411 of 432, 95.1% |
| Recommendations the rules found | 411 of 483, 85.1% |
| Business cards found | 304 of 304 |
| Recommendations outside business cards, found | 107 of 179, 59.8% |
The rules rarely invent a recommendation, but they miss some, mostly ones written into the prose. That has one consequence that travels with several figures below: the share of cited businesses that were also recommended is understated. In the hand-checked sample the understatement was about 9 to 10 points. A second check by an AI model (ChatGPT) agreed with the hand scoring on all 25 rows it was given, but those rows were mostly easy ones, so it supports typical cases, not the hard ones.
Two versions of every number
The registered protocol said to exclude businesses that Google itself did not corroborate: no Maps record, business card or Google reference anywhere in the data. The first version of our analysis script never applied that rule. We caught it while checking the findings. Applying it removes, among others, businesses whose only evidence is their own website, and those are disproportionately businesses that were cited and not recommended, which is the thing this study is about.
So every figure below is the registered version, with the rule applied as written, and where it matters the as-run version without it sits beside it. The protocol named the registered version as primary, and we kept it that way.
Most recommendations did not cite the business’s own website
| Recommendations with no citation to the business in the same answer | AI Overviews | AI Mode |
|---|---|---|
| Registered, own website only | 1,442 of 2,084, 69.2% (65.7% to 72.6%) | 1,932 of 2,120, 91.1% (88.4% to 93.4%) |
| As run, own website only | 1,618 of 2,293, 70.6% (67.3% to 73.6%) | 2,049 of 2,247, 91.2% (88.6% to 93.5%) |
| Registered, counting a Google-hosted reference to the business (all name matches) | 1,128 of 2,084, 54.1% | 722 of 2,120, 34.1% |
| Registered, same, strictest matching | 1,177 of 2,084, 56.5% | 906 of 2,120, 42.7% |
| As run, counting a Google-hosted reference to the business (all name matches) | 1,304 of 2,293, 56.9% | 827 of 2,247, 36.8% |
Recommendations are counted once per answer. Intervals are bootstrap intervals over the 200 questions. The Google-hosted rows are a sensitivity we did not register in advance; the strictest row leaves out references to places shown under more than one business and requires the displayed name to match the recommended name exactly. In the hand-checked sample, correcting the rules’ missed recommendations lowered the own-website shares by about 2 points.
The two surfaces recommended a similar number of businesses per answer and cited very different things. In commercial answers, AI Overviews recommended 3.21 businesses on average (644 answers) and AI Mode 3.02 (700 answers). AI Overviews answers carried 2.08 citations to business websites on average; AI Mode answers carried 0.72.
What AI Mode answers did list is visible in their sources as a whole. On commercial questions, 67.3% of AI Mode’s top-level references (1,598 of 2,375) pointed at Google’s own properties rather than any outside website; on informational questions it was 1.9% (11 of 579). For AI Overviews on commercial questions it was 17.9% (397 of 2,221). That is the clearest difference between the two surfaces in this study. It describes what the answers contained, not how Google built them.
Whether a cited business was recommended varied with the type of question
Turn it around and ask: when an answer cites a business’s website, does the same answer recommend that business?
| Cited businesses recommended in the same answer | AI Overviews | AI Mode |
|---|---|---|
| Registered | 642 of 933, 68.8% (62.9% to 73.7%) | 188 of 366, 51.4% (41.9% to 60.5%) |
| As run | 675 of 1,065, 63.4% (57.3% to 68.6%) | 198 of 480, 41.2% (33.0% to 49.7%) |
Counted as cited business-and-answer pairs. Intervals are bootstrap intervals over the 200 questions. Both rows are understated by the missed recommendations. In the hand-checked sample, correcting them moved AI Overviews from 62.8% to 72.3% (59 and 68 of 94) and AI Mode from 56.9% to 66.7% (29 and 34 of 51). Those are corrections within one sample, not a correction we can apply to the whole study, so we have not adjusted the figures above.
The averages hide the real pattern, which is the kind of question being answered.
| Type of question | AI Overviews | AI Mode |
|---|---|---|
| Commercial discovery | 245 of 276, 88.8% | 87 of 103, 84.5% |
| Recommendation | 343 of 417, 82.3% | 53 of 74, 71.6% |
| Comparison and trust | 46 of 125, 36.8% | 43 of 93, 46.2% |
| Informational | 8 of 115, 7.0% | 5 of 96, 5.2% |
Registered analysis; cited business-and-answer pairs; understated by missed recommendations.
Informational and comparison answers were less likely to recommend businesses they cited. Commercial discovery and recommendation answers showed much more overlap between the two. “Cited but not recommended” was far more common in some kinds of answer than others.
The same split accounts for most of the difference between the surfaces. Overall, AI Overviews recommended a cited business 17.4 points more often than AI Mode did (68.8% against 51.4%; bootstrap interval for the difference 8.9 to 26.2 points). Hold the type of question constant and the difference is no longer distinguishable from zero: the odds of a cited business being recommended on AI Mode, against AI Overviews, were 0.89 times as high (95% CI 0.56 to 1.42). AI Overviews cited more businesses inside commercial answers.
Recommendations changed from day to day
Before looking at what went with being recommended, it matters how much the recommendations moved.
The unit here is a business-and-question combination: one business, recommended at least once for one question. For each combination we counted how many of the rounds that produced an answer to that question also recommended that business.
| Business-and-question combinations recommended at least once | AI Overviews | AI Mode |
|---|---|---|
| Recommended in every round that produced an answer | 93 of 1,113, 8.4% | 59 of 1,109, 5.3% |
| Recommended in one round only | 601 of 1,113, 54.0% | 603 of 1,109, 54.4% |
More than half of the combinations recommended at least once were recommended in one round only. Fewer than 1 in 10 recurred in every round. Because our rules miss some recommendations, the every-round share is likely understated; we know the direction of that bias, not its size.
A single check of whether AI recommends you is one reading of a number that moves. Everything in the rest of this report is about patterns across that movement, not about any one answer.
Being cited went with being recommended, not against it
This is where the study did not go the way we framed it. The question behind it was whether AI search keeps “source of information” and “recommended provider” apart: whether a business could be the one an answer draws on without being the one it suggests.
On the data, the two went together.
- Of businesses cited on an informational question, 70.6% (48 of 68) were recommended on a commercial question in the same industry on AI Overviews, and 65.4% (34 of 52) on AI Mode.
- To test that against a fair baseline, we took the businesses in each industry’s Google Maps results, 535 business listings across the 20 industries, which are there whether or not AI picks them.
| Within the Google Maps results | AI Overviews | AI Mode |
|---|---|---|
| Recommended on a commercial question, if cited on an informational question | 16 of 21, 76.2% | 20 of 23, 87.0% |
| Recommended on a commercial question, all the others | 186 of 514, 36.2% | 197 of 512, 38.5% |
| Odds ratio, pooled across industries (Mantel-Haenszel) | 4.85 (95% CI 1.74 to 13.58) | 16.37 (95% CI 3.71 to 72.30) |
Registered analysis; business listings within an industry. Read those odds ratios with their intervals. Only 21 and 23 cited businesses sit inside the Maps results, so the true figure could be far smaller or larger than the middle value. No single industry showed a significant association on its own; the evidence is the association pooled across industries. It held when we used the Maps results alone without local-pack listings, and when we kept only businesses with a website. If missed recommendations fall mostly among businesses that were not cited, these odds ratios would be lower; we could not establish which way the bias runs.
Two things keep this in proportion:
- Informational citation was rare. Of the businesses recommended on a commercial question, only 6.6% (48 of 729, AI Overviews) and 5.0% (34 of 678, AI Mode) had ever been cited on an informational question in their industry. Most recommended businesses were never an information source at all.
- The separation does exist, for a minority. 20 of 68 (AI Overviews) and 18 of 52 (AI Mode) businesses cited on an informational question were never recommended on a commercial question in their industry. That is the kind of case that started this study. It is roughly 3 in 10 cited sources, not the rule.
So our working hypothesis, that being cited as a source and being recommended as a provider are largely separate, is not supported. The two are different measurements that, in this sample, went together. Nothing here shows that earning an informational citation makes a business more likely to be recommended.
Review count remained associated with recommendation in prose-only models
For the businesses in each industry’s Google Maps results, we modelled which ones were recommended in commercial answers, using their review count, star rating, Maps position and whether they ranked in Google’s organic top 10.
| Odds ratio per standard deviation (95% CI) | AI Overviews, all recommendations | AI Overviews, written in the answer only | AI Mode, all recommendations | AI Mode, written in the answer only |
|---|---|---|---|---|
| Review count (log) | 1.99 (1.74 to 2.27) | 1.43 (1.18 to 1.72) | 1.80 (1.58 to 2.06) | 1.63 (1.37 to 1.94) |
| Star rating | 1.68 (1.21 to 2.33) | 1.25 (0.87 to 1.80) | 1.56 (1.09 to 2.23) | 1.06 (0.78 to 1.43) |
| Maps position (a higher number is a lower position) | 0.77 (0.69 to 0.86) | 0.93 (0.79 to 1.09) | 0.93 (0.84 to 1.03) | 0.92 (0.78 to 1.09) |
Registered analysis; GEE logistic models clustered by question. Fitted on 15,791 (AI Overviews) and 17,185 (AI Mode) business-and-answer rows from 460 businesses and 644 and 700 commercial answers, after dropping 1,420 and 1,540 rows with no rating or review count. They cover only businesses with a Google Maps or local-pack record: 27.7% of the business-and-industry pairs AI Overviews recommended (202 of 729) and 32.0% of those AI Mode recommended (217 of 678).
The “written in the answer only” columns are a check on our own extraction. Business cards are detected perfectly and they display the star rating; recommendations written in the text are detected about 6 times in 10. Review count kept its association in the text-only models on both surfaces. For star rating the text-only results were inconclusive: both intervals include 1, which is not the same as showing there is no association. One possible reason, which we did not test, is that the rating association runs through the business card itself.
The simpler comparison agrees. As in September, the median star rating barely differed between businesses in the Maps results that were ever recommended and those that were not: 4.90 against 5.00 on AI Overviews (201 and 290 businesses) and 4.95 against 4.90 on AI Mode (216 and 275). Median review counts did differ: 70 against 21 on AI Overviews, 68 against 19 on AI Mode.
On AI Mode, appearing in the local pack was associated with recommendation too: odds ratio 5.77 (95% CI 4.19 to 7.93). We did not model the local pack on AI Overviews, because there a business card and a local-pack listing are partly the same observation.
Ranking organically went with being cited even more than with being recommended
The largest numbers in this study are also the least useful on their own, so they come last.
| Ranking in the organic top 10 for the same question and round | Odds ratio, being cited (95% CI) | Odds ratio, being recommended (95% CI) |
|---|---|---|
| AI Overviews | 103.35 (66.31 to 161.09) | 20.03 (13.60 to 29.49) |
| AI Mode | 11.17 (5.98 to 20.84) | 4.20 (2.88 to 6.11) |
Same models, same rows and the same coverage limit as the review table: businesses with a Google Maps or local-pack record only. Organic results were captured from the AI Overview results page.
On both surfaces, ranking organically was associated with both outcomes, and more strongly with being cited than with being recommended. On the log scale the gap was 1.61 (bootstrap interval 1.07 to 2.10) on AI Overviews and 0.92 (0.22 to 1.51) on AI Mode. It held when we kept only businesses with a website, which matters because a business needs a website both to rank and to be cited.
Part of the AI Overview figure is close to built in: an AI Overview and the organic results below it are on the same page, for the same search, so they will overlap whatever the mechanism. An odds ratio is not a probability and it is not a ranking-factor weight. None of this shows that ranking organically gets a business cited or recommended.
What the study did not find
- AI Mode did not recommend more businesses than AI Overviews. We predicted it would recommend more businesses with fewer website citations. The second half held; the first did not: 3.02 against 3.21 per commercial answer (bootstrap intervals 2.68 to 3.37 and 2.83 to 3.61, overlapping).
- Citation and recommendation were not separate (above).
- We cannot say which sources each part of an answer leans on. We tried to split citations by the section of the answer they supported: explanation or provider list. The section labels came from a word list we never validated, and our class for business websites was only 70% precise in a hand check (28 of 40). As classified, business websites were cited more in explanations than in provider lists, the opposite of our prediction. We are not reporting that as a finding.
- The SEO agency comparison was too small to test anything. For 15 SEO agencies we recorded what their websites and profiles showed. All 15 had an explicit service page, so there was no contrast to test. Spearman correlations with how often each agency was recommended: named people on the website 0.63, independent editorial mentions 0.27 (15 agencies each), Google review count 0.87 (only 5 agencies had one). With 15 agencies, these describe this group and nothing wider.
We are not publishing a ranking of the most-recommended businesses. Recommendation frequency in one week of one question set is not a measure of which provider is best.
How this compares with our September study
The Perth AI Search Study 2026 asked different questions, ran three rounds instead of five and used different extraction rules (September methodology). Every September figure below still reproduces from the September records. Because the question sets, definitions and dates all differ, none of these comparisons shows on its own that Google changed between September and November.
Google-hosted references in AI Mode. In September, 85.8% of AI Mode’s references (6,692 of 7,804) pointed at Google’s own properties. On the 20 questions asked in both studies the share was similar: 88.0% in September (805 of 915) and 86.5% in November (345 of 399). The number of references was not: on those same questions, an AI Mode answer carried about 15.2 references in September (60 answers) and 4.0 in November (100 answers), with the same request settings. We counted that from the raw responses Google returned, not from our extraction. This data does not tell us why the number fell.
How often recommendations recurred. In September we reported that 31% of recommended businesses on AI Mode and 11% on AI Overviews were recommended in all three rounds (1,063 of 3,421 and 23 of 219 business-and-question combinations, with no exclusions applied). This study’s 5.3% and 8.4% use five rounds, a harder bar to clear, so we also measured November on September’s three-round definition, applying the same exclusions to both studies and pooling November’s three consecutive three-day windows:
| Recommended in all three rounds | September | November, September’s definition |
|---|---|---|
| AI Mode | 1,050 of 3,347, 31.4% | 368 of 2,335, 15.8% |
| AI Overviews | 19 of 154, 12.3% (30 questions) | 328 of 1,703, 19.3% |
September counted Google places-panel listings in AI Mode as recommendations; counting them the same way in November gives 15.4% (399 of 2,587). On the 20 questions asked in both studies, the AI Mode figures were 33.8% (119 of 352, 14 questions) and 13.3% (70 of 525). The September AI Overview figure rests on 30 questions and 154 combinations, too few to read a direction from. The AI Mode figures are lower in November on every version of the comparison we ran; with different question sets and only two measurement periods, we do not read that as a trend.
How often an AI Overview appeared. An AI Overview appeared on 94.2% of our searches (942 of 1,000), against 65.5% in September (393 of 600). On the 20 questions asked in both studies the rates were close, 95.0% (57 of 60) and 97.0% (97 of 100), which is consistent with most of the gap coming from the different question sets. It does not rule out a change over time.
Organic overlap. As in September, businesses recommended by AI Overviews were far more often in the organic results for the same question than those recommended by AI Mode: 63.0% (529 of 840) against 30.0% (209 of 696) in November, and 66.8% (189 of 283) against 22.7% (428 of 1,887) in September. These count distinct question-surface-business recommendations with a known website; November uses commercial questions.
Concentration. Among commercial questions that produced at least one recommendation, no business was recommended on more than 9.3% of them (11 of 118 on AI Mode; 11 of 119, 9.2%, on AI Overviews). In September, on the same definition, the highest was 10.2%.
Business websites among AI Overview sources. On a strict definition, 44.0% of AI Overview top-level references on commercial questions went to a business’s own website (977 of 2,221), against 38.2% in September (1,086 of 2,845). We are not restating September’s broader share, because the two studies define a business website differently.
Nor do we restate September’s organic top 10 figure for businesses recommended by three or four AI systems: this study covers two. The organic finding above points the same way as September’s, and that is all we claim.
What went wrong
These are the mistakes we found in our own work while producing this study. Each one changed a number or a sentence before publication.
- The registered exclusion rule was never applied. The first analysis script left out the rule that excludes businesses Google did not corroborate. It changed the central measure: on AI Overviews the rule removes 132 of 1,065 cited pairs, 99 of them cited and not recommended. We found it while checking the findings, applied the rule as registered, and show both versions.
- Recurrence was counted against five rounds even when a question had fewer answers. That understated recurrence on AI Overviews: 5.3% instead of 8.4% (59 instead of 93 of 1,113 combinations). It is now counted against the rounds that produced an answer.
- The central test was first run on the wrong question. It first asked whether a business was cited anywhere, not whether it was cited on an informational question, as registered. We re-ran it as registered.
- Our first extraction rules in the pilot read labels as businesses. Version 1.0 counted section headings such as “Services” and “Rating”, trade words under explanatory headings, and links to source pages as recommended providers. In one answer it gave a listicle’s domain to a business the listicle mentioned. Four rule changes fixed this before the question set was frozen.
- We booked another tool’s spending as study cost. During the pilot, a drop in the shared account balance was first recorded as study spend. Other tools use the same account.
- We nearly published the wrong cost. Our planning notes for this report gave the cost as USD 2.46 for the five rounds and about USD 2.74 in all. Those were the amounts charged when the requests were sent, before DataForSEO refunded part of the charge on 99 AI Overview requests in the five rounds and 8 in the pilot. The costs below are after those refunds, summed from the cost field in every raw response, with the amounts charged at posting printed beside them.
- Our own website contradicted these findings. Until 14 November 2026, the AI SEO page on searchscope.com.au said that “citation means the AI is actively pointing users at you as the answer”. For informational questions, this data shows otherwise: cited businesses were recommended in the same answer 7.0% and 5.2% of the time. We corrected the page before publishing this report.
What we cannot tell you
- Why anything was cited or recommended. This is observational data. Every result is an association.
- Which source supported which recommendation. We recorded what appeared together in an answer, not what the answer relied on.
- Anything about logged-in or personalised results. These are logged-out, desktop, English results captured through an API, with Perth sent as the location.
- Anything long-term. Five consecutive days measure short-term variation.
- Exact recommendation levels. Our rules miss some recommendations, mostly those written in prose, so cited-and-recommended shares and recurrence are understated.
- Results for every business. The models behind the review and organic findings cover only businesses with a Google Maps or local-pack record.
- A clean separation between a recommendation and a Google Business Profile on AI Overviews. There, a business card is itself a Google-hosted listing, so recommendation and Google Business Profile presence are partly the same observation.
- Anything about ChatGPT, Gemini or Perplexity. They were not part of this measurement.
- Results outside these 20 Perth industries and these 200 questions.
What this means if you run a business
These follow from what we measured. None of them is a promise that a change will get you recommended.
- Track recommendations separately from citations. Of AI Mode’s business recommendations here, 91.1% had no citation to the business’s own website in the same answer, so a report that only counts citations of your site would have missed them.
- Count Google-hosted references to your business. Counting them cut the AI Mode share without any reference from 91.1% to between 34.1% and 42.7%. A visibility check that only looks for your domain will undercount how often answers reference you.
- Split your questions by type. Being cited on “how much does X cost” is a different result from being recommended on “who should I use”. Report them apart.
- Measure more than once. More than half of the business-and-question combinations recommended at least once in this study were recommended in one round only. One screenshot is not a position.
- Treat review count, organic rankings and local listings as signals worth examining. They were associated with recommendation here. That is not evidence that changing them changes AI answers.
Provenance
Published by: AI SEO Perth
Operated by: Search Scope (an SEO agency in Perth; SEO agency is a measured industry)
Researcher: Dorian Menard
Surfaces: Google AI Overviews, Google AI Mode
Collection period: 9 to 13 November 2026, 5 daily rounds from about 16:12 Perth
Questions: 200 across 20 industries, frozen 3 November 2026
Requests: 2,000 AI requests (1,942 answers) and 100 Google Maps captures
Extraction: frozen extractor v1.6, hand-checked on 195 answers
Analysis: registered protocol (primary); as-run beside it
Data source: DataForSEO API
Cost after refunds: USD 2.68 (five rounds USD 2.40; pilot USD 0.27; one keyword lookup USD 0.01)
Charged at posting: USD 2.74 (five rounds USD 2.46)DataForSEO charges when a request is sent and refunds part of the charge for some AI Overview requests. In the five rounds it refunded USD 0.0006 on 99 of 1,000 AI Overview requests: 58 that returned no AI Overview and 41 where the AI Overview came in the page itself rather than as a separate load. The cost after refunds is the sum of the cost field in every raw response.
This page exists as evidence, not because anyone searches for it. Search Console for searchscope.com.au, Search Scope’s own site, recorded 1,869 impressions and 0 clicks for “ai seo perth” between 1 August and 31 October 2026. That is interest in the topic, not in a study, and we are not claiming any demand for this report.
The observations behind this report are published alongside it.
The 2,000 AI observations are published as CSV and JSON, one row per request, with the businesses named and recommended as the answers wrote them and every top-level reference.
The full method, including the frozen 200-question set, definitions, extraction accuracy, the matching audit and the statistical methods, is on the methodology page.