Who Does AI Search Cite? 2026 Study of 372 Answers

Which websites do AI search engines cite when someone asks about Google Ads in German? To find out, I sent the same questions to three systems via their APIs once a month from July to October 2026: Perplexity (model sonar), OpenAI (gpt-4o-mini with web search) and Gemini (gemini-flash-latest with Google Search). The analysis covers 372 answers with 2,429 source mentions across 677 different domains. A mention means that a domain appears as a source in one answer.

In short, agency websites dominate, and which source gets cited depends heavily on which engine answers and in which month you ask. I recalculated every figure in this article from the raw data. You can download the domain table and the question list at the end.

Disclosure: mjmads.com is the publisher of this study and at the same time one of the measured domains. The questions come from my own visibility monitoring: commercial German-language Google Ads questions for which I want to be found myself. That is why mjmads.com does not appear in any ranking and is reported separately. Websites of other agencies are anonymised (Agency website 1, 2 and so on).
What was measured: queries through the application programming interfaces (APIs) with specific models, not the ChatGPT, Perplexity or Gemini apps. What users see in the apps may differ.
Data as of: runs on 16 July, 15 August, 11 September and 1 October 2026

Key findings at a glance

• Agency websites dominate: 67.3% of all source mentions, and as much as 82.8% for Gemini. This is partly down to the questions: 20 of the 34 explicitly ask for an agency or outside help.
• Google-owned pages account for 24.8% of mentions for OpenAI, but only 1.8% for Gemini.
• Many answers without evidence: 61.3% of OpenAI answers had not a single source, for Gemini it was 62.1%. Perplexity always cited sources.
• Long list, little concentration: 296 of the 677 domains (43.7%) were mentioned exactly once. The ten most cited domains (excluding mjmads.com) together account for 17.5% of all mentions.
• Little overlap: only 8.1% of sources per question and month were cited by more than one engine.
• October break: on the 22 questions asked in all four months, the agency share fell from 75.7% in September to 57.0% in October. Trade media rose from 7.1% to 15.3% in the same step.

Perplexity returns 15 sources per answer, OpenAI one

The first difference is obvious straight away: the three engines back up their answers to completely different extents.

• Perplexity (sonar)15.4
• Gemini (flash-latest)3.1
• OpenAI (gpt-4o-mini)1.1

Source: Own measurement, 372 answers, July to October 2026

124 answers per engine. Each domain counts once per answer. The APIs return sources in different formats, see Methodology.

Data as table

Engine	Average number of different sources per answer
Perplexity (sonar)	15.4
Gemini (flash-latest)	3.1
OpenAI (gpt-4o-mini)	1.1

Perplexity cited 15.4 different domains per answer on average, with a median of 16 and a maximum of 26. Not a single Perplexity answer came without a source. Gemini averaged 3.1 and OpenAI 1.1, and the median was zero for both. In other words, more than half of the answers from these two engines had no source at all. The maximum was 14 domains for Gemini and 8 for OpenAI.

Because Perplexity returns so many more sources, 1,913 of the 2,429 mentions (78.8%) come from Perplexity. All overall figures in this article are therefore heavily shaped by Perplexity. Where it makes a difference, I report the engines separately.

Why the numbers are not directly comparable

Part of the gap is technical rather than behavioural. The three APIs return sources differently, and the script reads a different field for each:

• Perplexity returns a list of search results for the answer. This is the most complete variant.
• OpenAI marks citations directly in the answer text (URL citations). According to the OpenAI documentation on web search, these inline citations show only the most relevant references. A separate list of all URLs consulted is often longer, according to the same page, and was not analysed in this measurement.
• Gemini returns grounding chunks, the supporting references from Google Search on which the answer is based.

So the numbers describe what each API reports as a source. Whether a model read more pages internally is not something the measurement can show.

Gemini has cited far fewer sources since August

More interesting than the average is the trend. For OpenAI, the share of answers without a source stayed roughly the same across all four months. For Gemini, it flipped in August.

• Perplexity
• OpenAI
• Gemini

0%25%50%75%100%

Jul 26Aug 26Sep 26Oct 26

Source: Own measurement, 372 answers, July to October 2026

July 22 questions per engine, from August 34 questions per engine

Data as table

Month	Perplexity	OpenAI	Gemini
Jul 26	0.0%	63.6%	18.2%
Aug 26	0.0%	61.8%	70.6%
Sep 26	0.0%	58.8%	70.6%
Oct 26	0.0%	61.8%	73.5%

In July, only 4 of 22 Gemini answers (18.2%) had no source. In October it was 25 of 34 (73.5%). Over the same period the average number of sources fell from 6.9 to 2.1. This is not caused by the new questions added from August onwards: on the 22 questions asked in all four months, the share without a source rose in exactly the same way, from 18.2% to 68.2% in August and 72.7% in October.

The data do not provide an explanation. One possible reason: the script queries the alias gemini-flash-latest. According to Google's documentation on Gemini models, such an alias points to the latest release of a model variant and is swapped with each new release. Which model was behind it on which run was not logged. So it is possible that a different model simply answered from August onwards.

It is also striking when OpenAI and Gemini cite sources at all. For questions asking for a provider, OpenAI had at least one source in 39 of 76 answers (51.3%), for knowledge questions only in 9 of 48 (18.8%). Gemini looks almost identical: 38 of 76 versus 9 of 48. Ask for an agency and you are more likely to get evidence than if you want to know how much budget Google Ads needs.

Two thirds of mentions go to agency websites

For the analysis I assigned each of the 677 domains to a type, based on its main purpose according to its homepage. The exact rule is in the Methodology section.

• Overall
• Perplexity
• OpenAI
• Gemini

• Agency websitesOverall: 67.3%Perplexity: 65.0%OpenAI: 55.5%Gemini: 82.8%
• Trade media & pressOverall: 7.6%Perplexity: 8.8%OpenAI: 0.7%Gemini: 4.2%
• Software & tool vendorsOverall: 7.1%Perplexity: 7.3%OpenAI: 12.4%Gemini: 4.5%
• Google-owned sitesOverall: 6.8%Perplexity: 6.5%OpenAI: 24.8%Gemini: 1.8%
• Comparison & directory portalsOverall: 5.4%Perplexity: 6.2%OpenAI: 2.2%Gemini: 2.9%
• mjmads.com (publisher)Overall: 1.9%Perplexity: 2.2%OpenAI: 0.7%Gemini: 0.5%
• OtherOverall: 1.4%Perplexity: 1.6%OpenAI: 0.7%Gemini: 1.1%
• Forums & socialOverall: 1.3%Perplexity: 1.6%OpenAI: 0.0%Gemini: 0.3%
• YouTubeOverall: 1.2%Perplexity: 0.9%OpenAI: 2.9%Gemini: 1.8%

Source: Own measurement, 2,429 mentions in 372 answers, July to October 2026

Base per engine: Perplexity 1,913, OpenAI 137, Gemini 379 mentions. Totals may differ from 100 due to rounding.

Data as table

Source type	Overall	Perplexity	OpenAI	Gemini
Agency websites	67.3%	65.0%	55.5%	82.8%
Trade media & press	7.6%	8.8%	0.7%	4.2%
Software & tool vendors	7.1%	7.3%	12.4%	4.5%
Google-owned sites	6.8%	6.5%	24.8%	1.8%
Comparison & directory portals	5.4%	6.2%	2.2%	2.9%
mjmads.com (publisher)	1.9%	2.2%	0.7%	0.5%
Other	1.4%	1.6%	0.7%	1.1%
Forums & social	1.3%	1.6%	0.0%	0.3%
YouTube	1.2%	0.9%	2.9%	1.8%

Across all engines, 67.3% of mentions go to agency websites. They are followed by trade media and press (7.6%), software and tool vendors (7.1%), Google-owned sites (6.8%) and comparison or directory portals (5.4%). Forums, social networks and YouTube barely matter, at just over one percent each.

The engines clearly set different priorities:

• Gemini relies almost entirely on agency websites (82.8%). Google-owned pages appear only seven times in its 124 answers, support.google.com and business.google.com never.
• OpenAI cites a quarter of its sources from Google (24.8%). The largest single item is google.com with 21 mentions, all of them for provider questions (21 answers to 10 questions). Which Google page exactly cannot be said, because the script only stores host names.
• Perplexity has the broadest mix. 98.4% of its answers contain at least one agency website, 57.3% at least one trade media outlet, 52.4% at least one Google page and 41.9% at least one comparison portal.

Knowledge questions are also backed by agency pages

That agency websites are so strongly represented is partly built into the questions. 20 of the 34 analysed questions explicitly ask for an agency, a freelancer or outside help, for example "Google Ads Agentur Köln Empfehlung" (Google Ads agency Cologne recommendation) or "Google Ads Konto gesperrt: wer kann helfen?" (Google Ads account suspended: who can help?). The other 14 are knowledge questions such as "Wie viel Budget braucht man für Google Ads?" (How much budget do you need for Google Ads?). Even when you separate the two groups, the picture remains clear.

• Provider questions
• Knowledge questions

• Agency websitesProvider questions: 73.2%Knowledge questions: 54.9%
• Trade media & pressProvider questions: 8.5%Knowledge questions: 5.9%
• Comparison & directory portalsProvider questions: 7.3%Knowledge questions: 1.5%
• Google-owned sitesProvider questions: 4.0%Knowledge questions: 12.6%
• Software & tool vendorsProvider questions: 3.0%Knowledge questions: 15.8%
• Remaining typesProvider questions: 4.1%Knowledge questions: 9.3%

Source: Own measurement, 372 answers, July to October 2026

Provider questions 1,644 mentions in 228 answers, knowledge questions 785 mentions in 144 answers. Remaining types: YouTube, forums & social, other, mjmads.com.

Data as table

Source type	Provider questions	Knowledge questions
Agency websites	73.2%	54.9%
Trade media & press	8.5%	5.9%
Comparison & directory portals	7.3%	1.5%
Google-owned sites	4.0%	12.6%
Software & tool vendors	3.0%	15.8%
Remaining types	4.1%	9.3%

For provider questions, 73.2% of mentions go to agency websites and comparison portals reach 7.3%. For knowledge questions the agency share drops, but at 54.9% it is still above half. Software vendors (15.8%) and Google-owned pages (12.6%) gain instead. 36.8% of answers to knowledge questions contain at least one Google page.

One limitation belongs here: the data contain only domains, not paths. Whether an agency's guide article or its homepage was cited, I cannot say. For knowledge questions a blog article seems likely, but it is not proven.

The most cited domains

The next figure shows which individual domains appear most often. mjmads.com is excluded here and reported separately further down.

• Agency website 163
• support.google.com62
• business.google.com47
• Agency website 247
• Agency website 346
• omr.com37
• Agency website 432
• Agency website 531
• agenturtipp.de30
• Agency website 630
• handelsblatt.com29
• youtube.com29
• Agency website 729
• sortlist.de28
• Agency website 826

Source: Own measurement, 2,429 mentions in 372 answers, July to October 2026

Number of answers in which the domain appears as a source, all engines and months. Agency websites anonymised and numbered by mentions. Ties in the order of the domain table.

Data as table

Domain	The 15 most cited domains (excluding mjmads.com)
Agency website 1	63
support.google.com	62
business.google.com	47
Agency website 2	47
Agency website 3	46
omr.com	37
Agency website 4	32
Agency website 5	31
agenturtipp.de	30
Agency website 6	30
handelsblatt.com	29
youtube.com	29
Agency website 7	29
sortlist.de	28
Agency website 8	26

At the top are one agency website with 63 mentions and support.google.com with 62. Google's help centre is the broadest source: it was cited for 20 of the 34 questions, business.google.com for 19. The top 15 include eight agency websites, two Google sites, two media outlets (omr.com, a German digital marketing media brand, and handelsblatt.com, a German business daily), two comparison portals (agenturtipp.de and sortlist.de) and YouTube.

More important than the top, though, is the breadth. The ten most cited domains together account for only 17.5% of all mentions. 296 domains, or 43.7%, were mentioned exactly once in four months. So the AI search engines draw their evidence from a very long field rather than from a handful of fixed addresses.

Each engine has its own ranking

Looking at the individual engines, there is hardly a shared top:

For Perplexity, two Google pages are among the top three. For Gemini, agency websites occupy the first five places, and the first other domain is youtube.com with 7 mentions, tied with five more agency websites. handelsblatt.com has 29 mentions, all from Perplexity. google.com, on the other hand, is never cited by Perplexity and 21 times by OpenAI. If you check only one engine, you get a rather different picture than someone who compares all three.

mjmads.com, reported separately

mjmads.com was cited as a source 45 times in the analysed answers (Perplexity 42, Gemini 2, OpenAI 1), for 22 of the 34 questions. That is 1.9% of all mentions. In the ranking, the domain would be in sixth place. By month: 9 mentions in July, 14 in August, 17 in September and 5 in October.

This figure says little about how visible mjmads.com is in general. The questions were selected for my own monitoring, and several of them cover topics I have published articles on myself, for example gold buying, funeral homes, jewellers and catering. I asked the two brand questions ("Wer ist MJ Marketing (mjmads.com)?" and "Google Ads Agentur Mijo Jurisic Erfahrungen") but excluded them from the analysis.

In October the source mix shifted

For a fair comparison over time I use the 22 questions that were asked in all four months. In July the question list consisted only of these 22 (plus the two brand questions). From August, 12 more were added.

• Agency websites
• Trade media
• Google
• Portals

0%25%50%75%100%

Jul 26Aug 26Sep 26Oct 26

Source: Own measurement, 22 identical questions, all three engines, July to October 2026

Base: July 473, August 436, September 449, October 391 mentions. Google = Google-owned sites, Portals = comparison and directory portals.

Data as table

Month	Agency websites	Trade media	Google	Portals
Jul 26	76.5%	4.7%	3.8%	3.8%
Aug 26	73.4%	5.5%	6.2%	6.2%
Sep 26	75.7%	7.1%	4.9%	5.6%
Oct 26	57.0%	15.3%	8.2%	9.7%

From July to September the agency share was stable between 73 and 77%. In October it fell to 57.0%. At the same time, the share of trade media doubled from 7.1% to 15.3%, and portals and Google pages also gained. The shift is driven by Perplexity, which supplied 321 of the 391 mentions for these 22 questions in October.

The ranking shows the same thing. Of the 20 most cited domains in August, 14 to 18 were still in the top 20 in September, but only 12 in October. The September range exists because seven domains share places 16 to 22 there. None of them disappeared completely: all 20 were cited at least once in October, they just lost places. Among those that dropped out were reddit.com (11 mentions in August, one in October) and several agency websites. New in the top 20 are herold.at (an Austrian business directory), developers.google.com, drweb.de, fuer-gruender.de and omt.de, plus three agency websites. Five domains were tied for places 18 to 22 in August, and none of them is in the October top 20, so the tie does not change the result of 12.

Turnover is also high across the whole period. Of all domains cited for the 22 identical questions in July or in October, only 29.9% appeared in both months. Of the ten most cited domains in July, five or six were still in the top 10 in October, because two domains share tenth place in July.

The break is clearest when you compare per question and engine: how many of the previous month's sources show up again for the same question?

• Perplexity
• OpenAI
• Gemini

• July to AugustPerplexity: 66.6%OpenAI: 20.7%Gemini: 17.2%
• August to SeptemberPerplexity: 69.0%OpenAI: 41.4%Gemini: 34.6%
• September to OctoberPerplexity: 40.2%OpenAI: 50.0%Gemini: 32.9%

Source: Own measurement, July to October 2026

July to August based on the 22 July questions, then 34 questions. Small base for OpenAI (29 to 44 sources per change) and Gemini (76 to 151).

Data as table

Month change	Perplexity	OpenAI	Gemini
July to August	66.6%	20.7%	17.2%
August to September	69.0%	41.4%	34.6%
September to October	40.2%	50.0%	32.9%

From July to August, Perplexity cited 66.6% of the previous month's sources again for the same question, and 69.0% from August to September. From September to October it was only 40.2%. For OpenAI and Gemini the bases are so small that I would not read anything into the fluctuations.

What happened in October cannot be read from the data. In terms of timing, a change at Perplexity fits: the Perplexity changelog has an entry for July 2026 saying "Sonar Chat Completions is now Agent API". The October run still used the previous endpoint and returned all 34 answers without errors. Whether Perplexity changed something behind it, I cannot prove. It is just as possible that the search results themselves changed.

The engines overlap very little

The last question: when three AI search engines receive the same question in the same month, do they cite the same sources? To answer it, I looked at every combination of question and month, 124 cases in total.

• All cases
• Only if all engines involved cite sources

• Perplexity and GeminiAll cases: 35.5%Only if all engines involved cite sources: 93.6%
• Perplexity and OpenAIAll cases: 19.4%Only if all engines involved cite sources: 50.0%
• OpenAI and GeminiAll cases: 11.3%Only if all engines involved cite sources: 40.0%
• All threeAll cases: 9.7%Only if all engines involved cite sources: 34.3%

Source: Own measurement, 124 combinations of question and month, July to October 2026

A case is one question in one month. Right bar: only cases in which every engine involved cited at least one source (Perplexity and Gemini 47, Perplexity and OpenAI 48, OpenAI and Gemini 35, all three 35).

Data as table

Engines	All cases	Only if all engines involved cite sources
Perplexity and Gemini	35.5%	93.6%
Perplexity and OpenAI	19.4%	50.0%
OpenAI and Gemini	11.3%	40.0%
All three	9.7%	34.3%

In 53 of 124 cases (42.7%), at least two engines cited a common domain, all three in only 12 cases (9.7%). This is partly because OpenAI and Gemini often cite no source at all. If you look only at cases where both sides had sources, a clear pattern emerges: Perplexity and Gemini shared at least one domain in 44 of 47 cases. So whatever Gemini cites almost always appears somewhere in Perplexity's long list too. OpenAI is much further away from both.

At the level of individual domains, the overlap is even smaller. Counting each domain once per question and month gives 2,235 sources. Only 181 of them (8.1%) were cited by at least two engines, and only 13 by all three. The shared sources are mostly agency websites (139 of 181). The most frequently shared was Agency website 1, in 14 cases.

The Gemini effect shows here as well: in July, at least two engines still shared a domain in 17 of 22 cases (77.3%). From August to October it was only 12 of 34 (35.3%) each month.

What you can take from the data

The sample is small and limited to one topic. The following points are therefore observations, not rules. To avoid any false expectations: I do not offer SEO. I am describing what the data show, not a service.

Check your own questions

Which sources get cited depends heavily on question, engine and month in this measurement. A single query in an app therefore tells you little. It makes more sense to keep a fixed list of the questions your customers really ask and to put them to several engines regularly. The question list in the downloads shows how such a list can be structured.

For provider questions, directories and media count too

If someone asks for an agency, they get comparison portals and trade media as evidence alongside agency websites. 23.2% of answers to provider questions contained at least one portal, 27.2% at least one trade media outlet. agenturtipp.de, sortlist.de and herold.at are among the frequently cited domains. Whether a directory listing leads to a mention, however, the measurement does not show. It only shows where the engines look for such questions.

For knowledge questions, Google itself is right there

For questions about how Google Ads works, Perplexity and OpenAI like to rely on Google's own help pages. support.google.com was cited for 20 of the 34 questions. A text that merely retells the help page competes directly with the original. What makes a page citable is not something these data answer. They only show who is already being cited.

One month is not a measurement

Between August and October, 8 of the 20 most cited domains lost their place, and in October Perplexity cited only 40.2% of the previous month's sources again. If you track AI visibility, compare months rather than reacting to a single run.

Paid visibility is a separate channel

You cannot buy a place in the source lists. Advertising in AI environments is a separate channel with its own rules. How Google serves ads around AI Overviews and AI Mode is covered in my article on Google Ads in AI search. How advertising in ChatGPT works is explained in the post on ChatGPT Ads in Germany, and there is a separate overview of advertising in AI assistants.

Methodology

The data come from a script I wrote for my own monitoring and run once a month. It sends each question individually to each engine, without a system instruction, without conversation history and without any location or language setting in the request, and stores the cited domains. If an error occurs, it tries the query a second time. No errors occurred for the analysed questions in any of the four runs. Answers without a source count as zero sources.

How the sources are captured

For Perplexity, the script reads the citations field and, if it is missing, the URLs from search_results. According to the Perplexity changelog, citations was removed in May 2025 and all applications should use search_results since then. I therefore assume that the sources come from search_results. Which field was filled was not logged by the script.

For OpenAI, the script reads the annotations in the answer text. According to the OpenAI documentation, a URL citation contains the URL, title and location of the cited source. The separate list of all sources consulted, which the same page says is often longer, was not requested. That explains part of the low source count.

For Gemini, the evidence sits in the answer's grounding metadata. According to Google's API reference, this is a list of supporting references retrieved from the specified grounding source, each with a title and URI. Because the URI can be a redirect link, the script uses the title if it looks like a domain name. In all four runs this was the case for every reference.

How the source types are defined

Each domain gets exactly one type, based on its main purpose according to its homepage. For 632 domains the homepage could be retrieved, for 45 it could not (bot protection, errors or no information), and there I classified by domain name or known purpose. The column "Basis of classification" in the domain table shows which applies to which domain.

Two borderline cases: I count omr.com as trade media even though it also runs a software review portal, because the editorial part dominates its homepage. Directories that also sell marketing (such as herold.at) count as portals, because business search is their main purpose.

Agency websites are numbered by their total number of mentions, Agency website 1 being the most cited. The numbers are identical in the text, figures and downloads. All other domains are named, but not linked.

The split into provider questions and knowledge questions follows a simple rule: any question containing one of the German words Agentur (agency), Freelancer, "wer hilft" (who helps), "wer kann helfen" (who can help) or "lassen" (have something done) counts as a provider question. All others are knowledge questions. As a result, "Google Ads selbst machen oder Agentur beauftragen?" (run Google Ads yourself or hire an agency?) counts as a provider question and "Google Ads für Wärmepumpen-Installateure" (Google Ads for heat pump installers) as a knowledge question.

The questions

All 36 questions in their German original with an English translation. I render the dash used in some original questions as a colon here. The CSV contains the exact wording as it was sent to the APIs.

Data downloads

Both files are free to use with attribution. The domain table contains all 677 domains with type, basis of classification and mentions per engine and month. The sum of all mentions gives the 2,429 used in this article.

Sources

• Perplexity, changelog (entries from May 2025 on the citations field and July 2026 on the Agent API): docs.perplexity.ai/changelog
• OpenAI, API documentation on web search (URL citations and source list): developers.openai.com/api/docs/guides/tools-web-search
• Google, Gemini API documentation on model versions and aliases: ai.google.dev/gemini-api/docs/models
• Google, Gemini API reference on generateContent and grounding metadata: ai.google.dev/api/generate-content
• Own measurement data: monthly runs on 16 July, 15 August, 11 September and 1 October 2026, published as a domain table and question list (see downloads)

Data as of: October 2026

Conclusion

AI search engines back up German Google Ads questions mainly with agency websites, complemented by Google's own help pages, trade media and comparison portals. How many sources an answer has, and which ones, depends so strongly on engine and month, though, that a single look into an app says very little. October showed how quickly the picture can shift. I plan to extend the study after further monthly runs. For context, my overview of search engine market share in Germany shows how large the AI chatbots' share of clicks to websites is so far.

If you want to test AI assistants as an advertising channel, I set up ChatGPT Ads from €1,500, and managing your Google Ads is available from €1,000 per month, with a 3-month start phase and cancellable monthly after that. And if you would like to automate monitoring like this for your own questions, that is a job for AI consulting, where the AI audit costs €700.