When does AI mention a business? Q3 2026 research

The findings at a glance

Source: Klusio’s Q3 2026 measurement archive, 1 July to 30 September 2026; publication version 1.3. The archive contains 18,206 valid answers to 1,538 questions covering 35 businesses. Five main AI systems are examined. These are observations about the measured questions and configurations, not a representative picture of all Dutch AI use.

Search queries add context: In the ChatGPT cost-question subset with visible search queries, a year was added in 105 of 128 answers (82%) and “Netherlands” in 102 of 128 (80%).
Place names are associated with more mentions: Giving each question equal weight, the tracked business was mentioned in 12.3% of answers to questions containing a place name versus 3.7% without one. Per answer, the figures were 9.7% and 2.9%. This does not prove that adding a place name causes a mention.
Cost questions mention the tracked business less often: The question-weighted figures were 1.7% for cost questions versus 6.9% for other questions; the answer-weighted figures were 1.7% and 5.0%. This concerns the tracked business, not whether the answer names any providers at all.
A mention does not reliably repeat: After an earlier mention, the repeat rate was 55.8% with equal weight per question and 67.6% per measured transition. The latter is based on 367 transitions, 82 questions and 22 businesses. A single answer is a poor basis for a fixed ranking.
How to read the weighting: Question weighting first calculates the share for each business and question, then averages those shares equally. Answer weighting gives frequently measured questions more influence. The reported Mantel–Haenszel ratios are separate within-group comparisons; their uncertainty intervals do not apply to the question-weighted averages.
Commercial interest: Klusio publishes this research and sells visibility software. This study does not measure the effect of using Klusio on growth. A mention, a citation, a recommendation and an enquiry are different outcomes.

1. ChatGPT expands the question before searching

The search-query analysis covers 1,267 ChatGPT answers to 719 questions for 34 businesses. Only answers with visible search queries can be analysed. ChatGPT used an average of 1.8 queries per answer; 51% contained more than one. Where the question explicitly named a place, 97% of the measured searches retained it. This describes supplied queries, not any additional search activity hidden from the archive.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Search additions: 1,267 ChatGPT answers, 719 questions, 34 businesses
Addition	Answers with addition / all answers	Per answer	Questions with addition / all questions	Equal weight per question
Adds “Netherlands”	471 / 1,267	37.2%	341 / 719	44.9%
Adds a year	329 / 1,267	26%	199 / 719	25.2%
Searches in English as well	248 / 1,267	19.6%	56 / 719	6.3%
Looks for official sources	243 / 1,267	19.2%	103 / 719	11.3%
Looks for reviews or experiences	116 / 1,267	9.2%	85 / 719	11.2%
Adds “best”, “top” or “reliable”	91 / 1,267	7.2%	24 / 719	2.7%
Looks for prices not requested in the question	81 / 1,267	6.4%	51 / 719	6.2%

Why weighting matters: English searches occurred in 248 of 1,267 answers (19.6%), but those answers came from just 56 of 719 questions. Giving each question equal weight produces 6.3%. “Best”, “top” and “reliable” additions likewise came from just 24 questions.
Counting individual search steps: There were 364 English queries among 2,249 individual search steps (16.2%). For brand questions, the corresponding count was 137 of 605 (22.6%). A percentage of search steps cannot be compared directly with the percentage of answers containing at least one such step.
Detection rules: Fixed rules and word lists identify the features; this is not a language-model classification. A feature counts only when present in a search query and absent from the original question. The rules recognise years from 2020 to 2039, Dutch references to the Netherlands, reviews and experiences, and predominantly English words from a fixed list.
A removed result: Version 1.0 included a 17% price-search figure for comparison questions. It was removed because that subset contained only 35 answers, below the publication threshold of 100.

Search additions by type of question

Cost questions often prompted a year and Dutch context; recommendation questions prompted reviews. The table gives both the answer counts and the number of distinct questions involved. Figures refer to this subset of the Q3 archive.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Question-type subsets of the visible ChatGPT search queries
Addition	Question type	Answer count	Per answer	Questions with addition / all in subset	Businesses
Adds “Netherlands”	Cost questions	102 / 128	79.7%	86 / 108	26
Adds a year	Cost questions	105 / 128	82%	89 / 108	26
Searches in English as well	Brand questions	81 / 228	35.5%	10 / 48	9
Looks for reviews or experiences	Recommendation questions	66 / 305	21.6%	54 / 202	32

What you can use: Check that your cost pages explain current prices, inclusions and exclusions, and any relevant Dutch conditions or rules. Change a date only when the content has genuinely been updated.
What remains untested: This analysis does not show that adding prices, a year or Dutch context makes a page more likely to be cited. That would require a separate test of pages and outcomes.
How Klusio uses this information: Where supplied by the system, Klusio shows the queries behind an answer. The Growth Engine can use them to prepare a more relevant page proposal. Your publication settings and approval still govern changes; the proposal is not proof of a causal SEO effect.

Four search examples, explained in English

These examples come from Klusio’s fixed question set on 29 and 30 September 2026. The translations below explain what the Dutch questions and queries mean. They are translations of the evidence, not separate measurements of English questions. The exact Dutch records follow.

How much does a new bathroom cost on average?: The searches became “new bathroom average cost [Netherlands] [2026]” and “bathroom renovation average cost [Netherlands] [2026]”.
Who is the best roofer in Utrecht?: The search became “best roofer Utrecht [reviews] roofer Utrecht”.
PVC or wooden window frames: which is better?: The queries compared maintenance, insulation and lifespan and added [Netherlands] and [Milieu Centraal], the Dutch consumer information organisation.
Cracks in my wall: is that dangerous?: The queries looked for structural cracks, foundations and damage to homes, adding [Netherlands] and [Rijksoverheid], the Dutch central government.

The original Dutch search records

The Dutch questions and queries are retained verbatim as research evidence. Square brackets mark additions highlighted in the original publication; they are not a claim that the user typed those words.

Wat kost een nieuwe badkamer gemiddeld?: nieuwe badkamer gemiddelde kosten [Nederland] [2026] badkamer verbouwen kosten gemiddeld [Nederland] [2026]
Wat is de beste dakdekker in Utrecht?: beste dakdekker Utrecht [reviews] dakdekker Utrecht
Kunststof of houten kozijnen: wat is beter?: kunststof houten kozijnen vergelijking onderhoud isolatie levensduur [Nederland] [Milieu Centraal] houten kunststof kozijnen onderhoud isolatie kozijnen [Nederland]
Scheuren in mijn muur, is dat gevaarlijk?: scheuren muur gevaarlijk constructief scheuren fundering [Nederland] [Rijksoverheid] scheuren muur fundering woning schade

2. When does AI mention the tracked business?

This analysis excludes questions that already contain the tracked business’s name: 15,123 answers to 1,524 questions for 35 businesses. Ratios compare within the same business and AI system; answer-source features also control for question type. A ratio above 1 means more mentions in the group with the feature; below 1 means fewer. An interval that includes 1 does not establish a clear difference.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Mention rates and within-group associations, Q3 2026
Feature	Question-weighted: with / without	Answer-weighted: with / without	Within-group ratio	95% interval, resampling questions	Businesses in same direction
Question names a place or region	12.3% / 3.7%	9.7% / 2.9%	5.08	3.15–11.18	16 of 21
Answer cites five or more sources	5.9% / 4.9%	5.6% / 3.2%	1.93	1.33–2.84	19 of 35
Question is longer than 12 words	13.4% / 4.2%	10.7% / 3.7%	1.78	1.21–3.06	15 of 25
Question asks for a recommendation	9.6% / 4.8%	6.8% / 3.7%	1.77	1.13–2.75	16 of 33
“Nearby”, without a place name	4.7% / 6%	4.6% / 4.5%	0.90	0.36–1.58	12 of 22
Comparison question	3.6% / 6.1%	1.8% / 4.6%	0.36	0.12–1.40	12 of 19
Cost question	1.7% / 6.9%	1.7% / 5%	0.25	0.12–0.44	20 of 30
Answer contains no citations	1.4% / 6.3%	0.6% / 5.5%	0.13	0.05–0.26	26 of 33

A place name is not the same as “nearby”: Place-name questions had a clearer association with a mention. For “nearby” without a place, the uncertainty interval crossed 1. That is uncertainty, not proof that location does not matter. Historical IP location, city, memory and session settings were not fully recorded.
The association remains within question types: For recommendation questions, the answer-weighted rates with and without a place were 15.1% and 3.6%; the within-group ratio was 3.7 (2.6–5.2). For cost questions, the rates were 3.1% and 1.2%, with a ratio of 2.4 (1.1–5.1). These are associations, not tests of local landing pages.
No citations does not mean no search: We can see which sources an answer cites. We cannot establish from a missing citation list whether a system searched in the background.
Comparison questions remain uncertain: Their question-level interval includes 1. The data does not support a firm claim that comparison questions reduce a business’s chance of being mentioned.
Practical use: Separate questions about costs and explanations from questions asking for a provider. Show genuine local projects, reviews and service areas where you actually work. Do not create near-identical pages that merely replace the place name. Whether local pages cause more AI mentions was not tested here.

Does the answer name other providers?

The tracked business appeared in 10.5% of answers naming other providers, versus 0.9% of answers without them. When providers were named, the list contained almost five on average. Local and recommendation questions named providers in more than half of measured answers. Explanatory and cost questions did so in roughly one in four.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Unbranded questions: other providers versus the tracked business
Question type	Answers	Names other providers	Tracked business, with others	Tracked business, without others	Questions	Businesses
All questions	15,123	37.9%	10.5%	0.9%	1,524	35
Local questions	2,152	65.1%	13.1%	0.5%	244	29
Recommendation questions	3,903	54.2%	12.4%	0.2%	372	35
Problem or emergency questions	301	56.5%	7.1%	0%	43	12
Comparison questions	650	31.4%	2.5%	1.6%	66	20
Cost questions	2,265	24.9%	5.7%	0.4%	283	30
Explanatory questions	5,852	21.7%	8.3%	1.3%	516	34

What the cost result means: 24.9% of cost answers named other providers. Even within that group, the tracked business appeared in only 5.7%. Do not interpret the lower overall mention rate as an absence of competing businesses.
Classification method: A retrospective gpt-4.1-mini analysis at temperature 0 read all 18,206 answers and recorded literal names in order, classifying them as local business, chain, platform, software, manufacturer, institution or other. A provider is a local business or chain. Known technology, software and platform brands are excluded even if classified as a chain.
Limits and sensitivity: The tracked business is recognised using its business name and website name. The classification is automated. Excluding Klusio’s own questions gives the same broad pattern: 11% with other providers versus 0.7% without. This study did not measure whether a place in a provider list leads to more enquiries.
How to use the distinction: For explanations and costs, useful information and source citations may be sensible goals. For local and recommendation questions, check whether you appear among the providers. Measure these types separately.

3. A mention is not a fixed position

A business mentioned once did not reliably appear in the next answer. With equal weight per question the repeat rate was 55.8%; per transition it was 67.6% (95% question-bootstrap interval 58.6–75.5%). There were 367 eligible transitions across 82 questions, 131 question-system series and 22 businesses. No separate interval was published for the 55.8% question-weighted figure.

Time between measurements: The median interval was five days; the 10th and 90th percentiles were one and eleven days. With at least seven days between measurements, 72% repeated, based on 105 prior mentions. Intervals varied, so these are not a fixed daily experiment.
A weekly sensitivity check: Selecting at most one measurement per question, system and week yielded a 68.5% repeat rate across 257 transitions. This retrospective restriction does not recreate measurements that were never taken.
A second AI system: The tracked business was also mentioned by a second system in 20.3% of 1,153 eligible comparisons (95% interval 16.2–25.0%). Excluding Klusio’s own questions gives 22.4%.
What to do with a one-off answer: Measure the same useful questions repeatedly and keep each system’s results separate. One successful answer is not a stable position or evidence of a lasting change. Whether a prospectively fixed measurement schedule changes these estimates remains to be tested.

The same question, the same day, different systems

The table compares the probability of a mention in each of two systems for the same question and day. The ratio is the first probability divided by the second, with a question-bootstrap interval. The Gemini–Claude interval crosses 1 and is not a clear difference.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Paired mention comparisons within the recorded Q3 configurations
Systems: first / second	Mention rates	Ratio	95% interval	Answer pairs
ChatGPT / Claude	7.9% / 3.5%	2.28	1.69–3.18	1,329
ChatGPT / Gemini	5.6% / 2.9%	1.97	1.39–2.98	2,345
Perplexity / Claude	6.5% / 3.5%	1.83	1.29–2.63	1,333
ChatGPT / Perplexity	5.8% / 4.1%	1.41	1.05–1.95	4,986
Perplexity / Gemini	3.9% / 2.9%	1.33	1.03–1.74	2,382
Gemini / Claude	4.6% / 3.6%	1.28	0.89–1.93	1,473

Not a universal league table: These results apply to the recorded access routes, model labels, questions and dates. Claude and Gemini had many empty results until around mid-August, so their comparisons lean towards the latter half of the quarter. Google AI results lacked enough businesses for the same comparison.
Different sources: The mean cited-domain overlap is shown below. Different systems can draw on different pages even when answering the same question. This alone does not establish which answer is better.

How much do cited sources overlap?

Each pair compares the unique domains cited by two systems for the same business, question and day. Jaccard overlap is the shared domains divided by all distinct domains cited by either system. For {A, B} and {B, C}, it is 1/3. The result is the mean of those ratios multiplied by 100; each pair has equal weight.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Mean Jaccard overlap of non-empty citation lists, Q3 2026
System pair	Mean overlap	Answer pairs
ChatGPT / Perplexity	6.5%	4,723
ChatGPT / Gemini	21.3%	1,647
ChatGPT / Claude	14.2%	726
Perplexity / Gemini	12%	1,697
Perplexity / Claude	14.4%	739
Gemini / Claude	16.1%	671

Which answers count: Both citation lists must be non-empty. A pair with an empty list is excluded, not given a zero. Questions naming a brand do count here.
A reproducibility limit: For duplicate records on the same day, the Q3 analysis retained the last non-empty source list read. The read order was not explicitly recorded. That limits reproduction of the selection for those duplicates.

Positions within a provider list

Among 609 eligible mentions covering 217 questions and 28 businesses, the tracked business was first in 43.5%, second in 17.1%, third in 13.5%, and fourth or later in 25.9%. These are positions among classified providers, not positions in the whole text or judgments of quality.

Stability, only when mentioned both times: Across 219 repeated pairs in which the business appeared both times, 49.3% retained exactly the same position and 80.4% moved by no more than one place. Of 168 pairs starting in the top three, 90.5% stayed there. This excludes cases where the business disappeared.
Exclusions: In 73 mentions (11%), the tracked business was not classified among the providers, for example because it was labelled software. Those mentions are excluded from the position analysis. Excluding Klusio’s own questions produces 45% first-place mentions and 91% top-three retention.
What remains unknown: A position only describes order. Whether being first produces more clicks or enquiries than a later position was not measured. Maintaining a mention and maintaining a rank should not be conflated.

4. The sources differ between the measured sectors

In the construction and maintenance group, 45% of answers cited a platform or comparison site. The corresponding figure for marketing and SEO was 5.1%. The groups contained 6,348 answers for 16 businesses and 7,832 answers for five businesses, respectively.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Share of answers citing each domain; domains can overlap
Measured group	Domain	Share of answers
Construction and maintenance	werkspot.nl	28%
Construction and maintenance	trustoo.nl	20%
Construction and maintenance	homedeal.nl	17%
Construction and maintenance	slimster.nl	14%
Construction and maintenance	bobex.nl	7.5%
Marketing and SEO	frankwatching.com	9.2%
Marketing and SEO	google.com	7.5%
Marketing and SEO	linkedin.com	7.1%
Marketing and SEO	searchlab.nl	6.4%

Look at the actual cited page: A platform citation may point to a general cost article, a business profile or a review page. If it contains information about your business, check that information. The research did not test whether buying a profile or collecting platform reviews causes more mentions.
These are two measured groups: Marketing includes just five businesses, with many questions from Klusio’s own set. Question types and systems also differ between the groups. These figures are not a sector-wide source ranking. Each percentage is a share of answers citing a domain, not market share.
What Klusio prepares: Klusio shows the cited sources for questions about your services and can prepare relevant mention opportunities. A mention only counts as implemented when it is actually online. Whether businesses with an active platform profile are mentioned more often within the same sector requires a separate test.

How this compares with other research

The original publication compared its findings with official documentation, academic papers and industry research, accessed on 2 October 2026. Much of that work concerns the US, and several publishers sell measurement software. The review did not find an equivalent study of Dutch AI answers at that time. Differences in dates, question categories and measurement units are stated below. This English edition translates that review; it is not a fresh verification of those sources.

Search queries

Comparison recorded for the Q3 2026 publication.

Searching the web with ChatGPT: OpenAI Help Center. Same broad direction. ChatGPT usually rewrites a question into one or more targeted searches. OpenAI’s example adds a year.
AI features and your website: Google Search Central · 10-12-2025. Same broad direction. AI Overviews and AI Mode can split a question into several related searches, known as query fan-out.
Patterns we see in ChatGPT query fanouts: Peec AI · 05-05-2026. Partial comparison. Peec reported an average of 2.1 queries per question and review searches without an explicit request. Our year-addition rate was higher, possibly because our group contained many cost questions.
ChatGPT searches in English, even when you don't: Peec AI · 12-02-2026. Partial comparison. Peec reported 43% English search steps for non-English questions, but did not include Dutch. Using the same unit, we counted 16.2%. Our 19.6% figure counts answers containing an English search and is not directly comparable.

Place names and location

Comparison recorded for the Q3 2026 publication.

Searching the web with ChatGPT: OpenAI Help Center. Same broad direction. OpenAI describes approximate IP-based location and advises specifying a city, neighbourhood or postcode when local results are unsuitable.
Web search: OpenAI API documentation. Same broad direction. The API web-search tool can use a supplied country, city, region and time zone. That capability does not establish which location settings were used in our historical measurements.
What 200k local AI searches tell us: BrightLocal · 16-09-2026. Partial comparison. This US study used place names in questions and found that ChatGPT handled distance more broadly than Google. It is only a partial comparison with Dutch questions.

Question type

Comparison recorded for the Q3 2026 publication.

Why 62% of AI citations don't lead to brand mentions: Semrush · 09-06-2026. Same broad direction. Informational questions could have citations without brand names in the answer. Semrush grouped “best” and “A vs B” together; our comparisons concern methods and materials, so the categories differ.
How query intent shapes brand competition across AI search engines: BrightEdge · 20-11-2025. Partial comparison. Consideration-stage questions such as “best …” produced the most brand names. These were US e-commerce questions, not the same group as our Dutch service questions.

Answers without citations

Comparison recorded for the Q3 2026 publication.

Global is Good, Local is Bad? Understanding Brand Bias in LLMs: Kamruzzaman et al., EMNLP 2024 · 20-06-2024. Same broad direction. The paper found a bias towards global over local brands in language models. That is consistent with the pattern in our answers without citations, but missing citations do not prove that no search occurred.
Generative Engine Optimization: How to Dominate AI Search: Chen et al., University of Toronto · 10-09-2025. Different finding. This paper found a strong preference for third-party sources over a brand’s own website. It studied US product brands. Our local service questions included the business’s own website more often, so the findings differ in context.

Changing mentions

Comparison recorded for the Q3 2026 publication.

AIs are highly inconsistent when recommending brands or products: SparkToro · 28-01-2026. Same broad direction. Repeated questions produced the same brand list less than once in a hundred attempts. A one-off list is a limited basis for a stable ranking. API measurements and individual user situations are not interchangeable.
What 200k local AI searches tell us: BrightLocal · 16-09-2026. Same broad direction. ChatGPT and Google AI shared around 19–20% of named businesses. A previously named business returned in roughly half of subsequent attempts. We measured 67.6% per transition and 55.8% with equal weight per question; the studies have different groups and methods.
86% of top mentioned sources are not shared across ChatGPT, Perplexity, and AI Overviews: Ahrefs · 12-06-2025. Same broad direction. Only seven of the fifty most-cited domains appeared across all three systems in that study.
Do AI engines cite the same sources?: Writesonic · 22-07-2026. Same broad direction. Writesonic reported low ChatGPT–Perplexity source overlap using Jaccard (0.13). Our mean was 0.065, or 6.5%. The questions, groups and periods differ, so this is not a direct benchmark.

Platforms as sources

Comparison recorded for the Q3 2026 publication.

The LLMs have spoken: Yelp is the top source for AI local search: Foundation en AirOps · 30-07-2026. Partial comparison. The US study found frequent citations of Yelp, BBB and Angi for local services. The broad pattern resembles our Dutch group, with different platforms.
Top sources and directories for local AI search: BrightLocal · 26-08-2026. Partial comparison. Google Business Profiles and Yelp were leading sources in the US, UK and Australia, with specialist directories varying by sector. This does not establish the same ordering in the Netherlands.

What you can do this month

Use the observations to decide what to inspect and measure. They do not justify promises of higher rankings, more mentions or more enquiries after a particular change.

Track different customer questions: A cost page can be useful as a cited source, landing page and decision aid even when the business name is absent from the answer. Measure that separately from provider recommendations.
Show what you actually do in each area: Use genuine projects, reviews and service-area information. Avoid copied pages with different place names. The question-weighted place-name association exceeded threefold, but the study did not test whether local pages contributed to it.
Review cost information: Check prices, inclusions, exclusions and relevant Dutch rules. A current year belongs on genuinely updated information, not on unchanged text.
Make genuine customer experiences findable: ChatGPT independently looked for reviews in just over one in five recommendation answers in the measured subset. This is a reason to inspect what it finds, not a promise that a review will be used.
Measure each system more than once: ChatGPT’s answer does not stand in for Gemini’s or Perplexity’s. Keep the question, date, system and evidence alongside the result.

Receive the next quarterly report

One email per quarter, with the method and change log. Quarters are only directly comparable once differences in the measured group, settings and cadence have been checked. Confirm your email address to subscribe.

Method and limitations

The measurements cover 1 July to 30 September 2026. Each valid answer is a separate observation; answers were not reused. The archive records NL / nl for all 18,206 valid records. The configurations below explain the gaps in historical settings.

The measured questions: The total includes 1,409 questions tracked for 35 small and medium-sized businesses, mainly in construction, maintenance, installation, business services and marketing; 97 questions from Klusio’s own set, producing 6,672 answers; and 32 fixed questions across eight sectors, started in late September.
Access routes: ChatGPT and Gemini used web-interface observations; Perplexity and Claude used APIs; Google used search results. Older routes are partly unknown. Google includes AI Overviews and AI Mode. One valid Copilot answer is in the archive total, but it is insufficient for a separate system-level finding.
An adaptive schedule: Questions were chosen around each business’s services and area. A changing answer was measured more often than a stable one. A question-system series had between 1 and 54 measurements, median 1; intervals were usually one to eleven days. This can bias answer-weighted comparisons.
The intended definition of a mention: The business name appears in answer text. A name only in a source URL is not a text mention. Questions naming the business are excluded from mention associations and repeatability, but retained for search-query analysis and source overlap. Automatic recognition sometimes differs from this intended definition; the diagnostic check is reported in full below.
Comparison and uncertainty: Mantel–Haenszel ratios compare within business and system. Source-related features also account for question type. The main 95% intervals resample whole questions 1,000 times to respect repeats. A separate 200-resample business check and a comparison excluding Klusio are also provided.
Publication thresholds: Substantive results require at least five businesses and one hundred answers; the fixed question set requires fifty answers. Smaller diagnostic counts describe data quality rather than market or system performance.
What this study cannot establish: It is not a representative sample of all Dutch AI questions, a randomised test of website changes, or a measure of recommendations and enquiries. Logged-in personalisation, memory and session histories were not fully recorded. Systems changed during the quarter.
Privacy and reproduction: Customer names, domains and individual questions are not published. The examples are from Klusio’s own fixed set. Public platforms and cited domains are named. Protecting the underlying answers limits independent reproduction of classifications and bootstrap calculations.

Technical appendix: all statistical comparisons

Basis: 15,123 answers to 1,524 unbranded questions for 35 businesses. The full table retains the original raw ratios, within-group ratios, three intervals and sensitivity checks. “With / without” means with or without the row’s feature. Scroll the table horizontally where needed.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Complete statistical appendix from publication version 1.3
Feature	Answer-weighted with / without	Raw ratio	Within business × system	95% per answer	95% per question	95% per business	Excluding Klusio	Within question type (95%)	Question-weighted with / without; ratio	Weekly selection ratio	Questions with feature
Question names a place or region	9.7% / 2.9%	3.39	5.08	4.19–6.16	3.15–11.18	2.71–19.33	5.08	3.94 (2.04–11.57)	12.3% / 3.7%; ratio 3.33	3.10	397
Answer cites five or more sources	5.6% / 3.2%	1.75	1.93	1.59–2.34	1.33–2.84	1.53–2.53	2.02	Not applicable	5.9% / 4.9%; ratio 1.20	1.71	1,522
Question is longer than 12 words	10.7% / 3.7%	2.87	1.78	1.48–2.14	1.21–3.06	1.16–3.81	1.77	2.01 (1.12–4.45)	13.4% / 4.2%; ratio 3.20	2.82	289
Question asks for a recommendation	6.8% / 3.7%	1.84	1.77	1.51–2.07	1.13–2.75	1.04–2.70	2.16	Not applicable	9.6% / 4.8%; ratio 2.01	1.88	372
“Nearby”, without a place name	4.6% / 4.5%	1.01	0.90	0.68–1.19	0.36–1.58	0.32–2.08	0.90	0.47 (0.11–1.05)	4.7% / 6%; ratio 0.77	0.85	111
Comparison question	1.8% / 4.6%	0.40	0.36	0.16–0.83	0.12–1.40	0.13–1.59	0.36	Not applicable	3.6% / 6.1%; ratio 0.59	0.39	66
Cost question	1.7% / 5%	0.33	0.25	0.17–0.35	0.12–0.44	0.08–0.60	0.25	Not applicable	1.7% / 6.9%; ratio 0.24	0.27	283
Answer contains no citations	0.6% / 5.5%	0.10	0.13	0.08–0.23	0.05–0.26	0.05–0.23	0.12	Not applicable	1.4% / 6.3%; ratio 0.23	0.12	843

Raw and stratified ratios answer different questions: For place names, the raw ratio is 3.39, compared with 5.08 within business and system and 3.94 within question type. Different groups have different baseline mention rates. The association therefore cannot be explained solely by place questions being more often recommendation questions. This is still not causal evidence.
Three uncertainty intervals: Treating each answer as independent produces an interval that is too narrow for repeated questions. The report therefore uses 1,000 whole-question resamples, a fixed seed and the 2.5th and 97.5th percentiles. For comparison questions, 797 draws were usable: some draws lacked an eligible comparison within a business and system. Resampling whole businesses 200 times is a further, stricter check.
Excluding Klusio: Klusio’s own questions produced more than a third of all valid answers. Removing them generally preserves or strengthens these associations. The place-name estimate is unchanged because Klusio’s own questions contained no place name.
Question weighting and cadence: The archive contains 5,660 question-system series. The 103 series with changing mention outcomes averaged 9.5 measurements, versus 2.6 for series never mentioning the business; series with at least one mention averaged 3.6. The median series had one measurement, the 90th percentile seven and the maximum 54. Cadence is therefore related to the outcome.
What changes under equal weighting: Question-weighted ratios remain 3.33 for place names, 2.01 for recommendations, 3.20 for long questions, 0.24 for costs and 0.23 for no citations. Five or more sources drops to 1.20. Comparison and “nearby” findings remain uncertain. These point estimates are not the Mantel–Haenszel ratios and do not inherit their intervals.
The weekly selection: At most one measurement per question, system and week retains 11,268 of 15,123 answers. The table’s weekly column shows the resulting ratios; repeatability is 68.5% over 257 transitions. This was a retrospective sensitivity analysis, not a pre-scheduled study.
Download definitions: The aggregate download includes published counts and additional definitions. Empty fields mean unavailable or not applicable, never zero. Equal-weighted means and adjusted ratios cannot be represented by a single integer numerator and denominator.

Place-name association by system

These are within-business comparisons for the recorded system configurations. Mention rates are answer-weighted; the within-business ratio need not equal the simple ratio of those rounded rates. The separate recommendation and cost comparisons above use answer-level uncertainty intervals.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Place names: ratios within business, Q3 2026
System	Within-business ratio	With a place name	Without a place name
ChatGPT	5.60	11.1%	3.9%
Perplexity	4.00	8%	2.8%
Gemini	5.70	12.1%	1.5%
Claude	6.90	8%	2.1%

Recorded configurations and availability

The following counts were checked directly against the Q3 archive on 4 October 2026, without changing answers or classifications. Valid records total 18,206. Empty, error and unavailable records are excluded from valid answers. These are archived statuses, not a complete failure rate for every call ever started.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Routes, stored model labels and archived statuses
System	Route	Stored labels for valid answers	Valid measurement dates	Valid	Empty	Error	Unavailable
ChatGPT	Web interface; older route partly unknown	gpt-5-6: 579; gpt-4.1-mini: 11; scraper label without model version: 33; model not recorded: 5,273	2026-07-01 to 2026-09-30	5,896	0	69	0
Perplexity	API; older route partly unknown	sonar: 1,885; model not recorded: 4,061	2026-07-01 to 2026-09-30	5,946	0	56	0
Gemini	Web interface; older route partly unknown	3.5 Flash-Lite: 1,310; gemini-3.5-flash: 11; scraper label without model version: 33; model not recorded: 1,922	2026-07-07 to 2026-09-30	3,276	1,093	53	0
Claude	API; older route partly unknown	claude-sonnet-4-5: 44; claude-sonnet-4-5-20250929: 139; model not recorded: 1,406	2026-07-07 to 2026-09-30	1,589	1,093	0	0
Google AI Overviews	Google search results	Underlying model unknown; google-ai-overview is a product label	2026-07-07 to 2026-09-30	1,394	29	322	45
Google AI Mode	Google search results	Underlying model unknown; google-ai-mode is a product label	2026-09-29 to 2026-09-30	104	1	0	0
Copilot	Measurement route insufficiently documented	Model not recorded; too few usable answers for a system-level finding	2026-09-28 to 2026-09-28	1	2,486	0	0

Search settings, instructions and location

The historical records do not reproduce every aspect of a person’s AI session. Current software settings cannot establish what was active for every Q3 call.

Country and language: All 18,206 valid records store NL / nl. Some error records lack these fields.
Model labels: The configuration table translates the descriptions but preserves stored model labels. A scraper or product label is not a model version, and a label does not guarantee that a model stayed unchanged throughout the quarter.
Web search: Queries are available only when supplied. A complete historical search-mode and search-context configuration was not recorded for every call. Missing citations cannot establish whether background searching occurred.
Generation instructions: Historical instructions, temperature and other generation settings are not reproducibly stored for every route. The temperature-0 setting belongs to the later provider-classification analysis, not necessarily to the original answers.
Location and session: City, postcode, outgoing IP address, account settings, memory and session history were not recorded in full. OpenAI’s cited documentation describes approximate IP location and, depending on settings, memory, and a separate API location field. These capabilities explain why standardised measurements can differ from personal answers; they do not prove the settings of our archived calls.

What the name-recognition check shows

An additional literal-name search checked 1,200 answers. It searched business and website names, normalising spaces and case. This is an automated countercheck, not independent human adjudication; the second method can also miss or misidentify a name. The Q3 findings have not been corrected using this check.

Scroll horizontally to see every column. You can also focus the table and use the arrow keys.

Negative-record countercheck: diagnostic sample only
System	Records checked as negative	Name nevertheless found in text
ChatGPT	275	3
Gemini	145	5
Claude	89	1
Perplexity	234	0
Google AI Overviews	53	0
Google AI Mode	4	0

400 records classified as mentioned: The countercheck found the name in visible text in 391 records. In nine it appeared only in a link; all nine were Google AI Overviews records.
800 records classified as not mentioned: The countercheck found nine names in visible text: five Gemini, three ChatGPT and one Claude. A further three had a name only in a link and do not count under the text-mention definition.
How the sample was drawn: Using seed 20261004, 400 positive and 800 negative records were sampled separately from the base group. This is stratified, not a random cross-section with the original positive-to-negative proportions.
Small diagnostics, not a system ranking: The nine missed text mentions involved five local, two explanatory, one cost and one comparison question. The nine positive link-only cases involved eight explanatory questions and one cost question. These counts are too small for reliable error rates by question type.

Sensitivity to recognition differences

The base group contains 682 of 15,123 answers classified as a mention, about 4.5%. If both countercheck fractions (391/400 and 9/800) applied to the entire group, an illustrative calculation gives approximately 5.5%. This is a conditional scenario, not a new measured percentage, a validated correction or an uncertainty interval.

Why the same error rate cannot be assumed: Link-only positives were concentrated in one system. Applying the same rate to every question type, system ratio or repeatability result would be unjustified.
What remains clear: Most answers still lack the tracked name in this scenario. Whether precise differences between groups change has not been established. Search-query analysis and source-domain overlap do not use the tracked-name score and would not change solely from correcting it. A validated recalculation requires adjudicating disagreements and sufficient checks within each group.

Conditional illustration only

(682 / 15,123 × 391 / 400) + ((1 − 682 / 15,123) × 9 / 800) ≈ 5.5%

Comparing Q3 with future quarters

Q3 combines adaptive customer monitoring, Klusio’s own questions and a small fixed set begun in late September. Version 1.1 added question weighting and a name countercheck. Version 1.2 classified providers and positions retrospectively and imposed a weekly selection on the existing archive. None of those retrospective steps creates a prospectively fixed study.

A separate fixed series: Future quarter comparisons need a fixed question set, a fixed cadence and recorded configurations, separately from adaptive customer monitoring. Participating questions, model changes, missing observations and replacements must be documented each quarter.
Do not call every difference a trend: A weekly selection cannot restore missing observations. Changes relative to Q3 should only be called trends after checking these differences.
Reproduction limits: The aggregate CSV provides counts, percentages, weighting and definitions. It does not expose protected answers, so it cannot independently reproduce all bootstraps or recognition decisions. Where no whole-number numerator and denominator exist, those cells remain empty.

Three findings to share

You may reproduce these charts with attribution and a link to this report. Each chart includes the measured group, weighting and its main limitation. The values are read from the same public snapshot as the Dutch edition.

English searches: weighting changes the result: 6.3% / 19.6% — 1,267 answers · 719 questions · 34 businesses. Only ChatGPT answers with visible search queries.

Mentions with and without a place name: 12.3% / 3.7% — 15,123 answers · 1,524 questions · 35 businesses. Equal weight per question. Association, not causation.

Mentioned again after a previous mention: 55.8% / 67.6% — 367 transitions · 82 questions · 22 businesses. Only after a prior mention. Measurement intervals vary.

Version history

Every revision to the source report receives a version and explanation. “Editorial” means the measurements are unchanged and the wording is more precise. This English edition preserves the original measurement version and records its translation separately.

Version 1.3 · 4 October 2026

Changes recorded in the original publication.

Editorial: Question-weighted findings lead, with answer-weighted results alongside. Existing Q3 measurements are retained.
Correction: Over-broad statements about cost questions and “at most one in ten” are replaced by statements about the tracked business and measured group.
Addition: Configurations and archive statuses are checked. Unknown historical settings remain unknown. System comparisons apply to the tested configurations.
Addition: Name checking is described as an automated countercheck, with a conditional sensitivity calculation, not a validated correction.
Addition: Equal-weighted findings and definitions are added to the CSV, plus three reusable charts with group, unit and limitation.
Editorial: Jaccard denominators, source selection, the current citation and the distinction between retrospective Q3 analyses and future quarterly comparisons are clarified.

Version 1.2 · 4 October 2026

Changes recorded in the original publication.

Addition: All 18,206 answers are analysed to establish whether other providers are named and how often the tracked business appears alongside them.
Addition: Provider-list positions and their stability are analysed.
Addition: A check selects at most one measurement per question, system and week. The main associations remain.
Addition: All publishable aggregate counts become downloadable as CSV.
Editorial: The earlier “not separately measured” caveat about cost questions is replaced by the new provider analysis.

Version 1.1 · 4 October 2026

Changes recorded in the original publication.

Editorial: “There is no fixed AI top three” becomes “A mention is not a fixed position”: that original analysis measured mentions, not positions.
Editorial: The “nearby” heading is narrowed. No clear measured difference is not proof that location has no effect.
Editorial: The distinction between the tracked business and any providers in an answer is clarified.
Editorial: “Without sources” becomes “without citations”. Background search is unknown.
Correction: The 17% price-search result for comparison questions is removed because it rested on only 35 answers.
Addition: Search features receive numerators, denominators, distinct-question counts and detection rules.
Addition: English is also counted per search step (16%), matching the unit in Peec’s research.
Addition: Equal question weighting checks cadence bias. Place, recommendation, cost and citation findings remain; the five-source association weakens. Repeatability is 68% per transition and 56% per question, rounded.
Addition: Name recognition is checked on a 1,200-answer sample.
Addition: Each part explains how Klusio uses the observations for a business.

Version 1.0 · 2 October 2026

Changes recorded in the original publication.

Editorial: First publication.

Publisher, contact and citation

Klusio published the report and performed the automated analyses. Sicco Dijkman, founder of Klusio, is the contact for methodological questions. No independent external review of the underlying answers and calculations is recorded. This research is not a study of Klusio’s effect on customer growth.

Citation for this English edition

Klusio (2026). When does AI mention a business? AI research report Q3 2026, version 1.3, source updated 4 October 2026. English edition prepared 21 October 2026. https://klusio.nl/en/ai-research/q3-2026/

Apply the questions to your own business

Klusio tracks questions about your services and service area, shows the businesses and sources in each answer, and prepares relevant website improvements through the Growth Engine. Use your own evidence to decide which changes are worth making.