Foundations

GEO by search intent: which media AI answers cite

AI assistants do not cite the web the way Google ranks it. A University of Toronto study shows which kinds of sources they draw on for informational, comparison and purchase questions. What that means for owned, earned and paid media, and what communication teams should watch.

, 9 min read, 2 sources

In short

  • In the study’s pooled figure for 300 consumer-electronics questions, independent media and review outlets supply about half of all cited sources, brand-owned sites about a third and social platforms about an eighth.
  • The mix moves with intent: comparison questions are answered mostly from earned media, purchase questions mostly from brand sites.
  • Earned coverage decides whether a brand makes the shortlist; owned pages decide whether the answer at the moment of purchase is right. Both have to be measured per engine and per market.

1AI answers draw on a different web

Generative engine optimisation (GEO) asks a simple question: when someone asks an AI assistant, which sources does the answer rest on, and is your brand among them? A team at the University of Toronto measured how far that differs from classic search. Across 1,000 ranking-style questions in ten consumer topics, the domains cited by the AI systems overlapped with Google’s top ten results by only 4.0 % on average for GPT-4o and 15.2 % for Perplexity Sonar Pro, with Gemini 2.5 Flash (11.1 %) and Claude 4.5 Sonnet (12.6 %) in between. For GPT-4o the median overlap was 0.0 %: for more than half of the questions it cited no domain from Google’s top ten.(Chen et al., University of Toronto, 2026, p. 2)

Rank in Google is a weak guide to citation in an AI engine, since overlap with Google's top ten ranged from 4.0% to 15.2%.
Everything-PR, in its reading of the study, Everything-PR, June 2026, What does this mean for communications teams?

Overlap says that the sources differ. The more useful part of the study for communication teams is what kind of sources the engines cite, and how that changes with what the person asking wants to do.

2How the study sorts sources: Brand, Earned, Social

For this part the authors took 300 questions about consumer electronics, 100 each for three intents: informational (“How do OLED TVs work?”), consideration (“Best budget noise-canceling headphones under $200”) and transactional (“Buy Apple AirPods Pro 2 near me”). Every source cited by Google, GPT-4o, Claude 4.5 Sonnet, Perplexity Sonar Pro and Gemini 2.5 Flash was labelled as one of three types by GPT-4o, with links from known social platforms assigned to Social automatically and a manual spot check of a random subset.(Chen et al., University of Toronto, 2026, p. 3)

Brand refers to official company-owned domains (e.g., apple.com), Earned to independent media and review outlets (e.g., forbes.com), and Social to community or user-generated platforms (e.g., reddit.com).
Chen et al., University of Toronto, 2026, p. 3

These three types are not the owned, earned and paid media communication teams plan in: the study sorts by the site a source is on, not by how the text came about. How the two schemes relate is set out in the note on owned, earned and paid at the end.

3What AI answers cite, by intent

Source types in cited sources, by search intent

Share of cited sources in %, row “All Models” of fig. 2, pooled as printed; 300 consumer-electronics questions (100 per intent); categories as labelled in the study

  • Earned
  • Brand (owned)
  • Social

Informational · learning

Earned
48 %
Brand (owned)
40 %
Social
12 %

Consideration · comparing

Earned
66.3 %
Brand (owned)
17.1 %
Social
16.6 %

Transactional · buying

Earned
38.3 %
Brand (owned)
52.9 %
Social
8.8 %

All intents

Earned
51.2 %
Brand (owned)
36.3 %
Social
12.5 %

The leading source type changes with intent: earned media while people learn and compare, the brand’s own site when they are ready to buy.

Source: Chen et al., University of Toronto, 2026, p. 4, fig. 2

Informational: earned slightly ahead of owned

For questions that explain how something works, earned sources supply 48.0 % of citations and brand sites 40.0 %. The authors note that the AI systems vary widely here in how much weight they give earned versus brand sources. For a brand this is the stage where its own explanatory content still has a real share of the answer, provided it explains rather than sells.(Chen et al., University of Toronto, 2026, pp. 3-4)

Consideration: the answer is written from earned media

When people compare (“best … under $200”), earned sources rise to 66.3 % and brand sites fall to 17.1 %. Per AI system, earned media supplied 59 % to 86 % of citations, while Google’s results for the same questions were led by social sources (41 %). This is the stage where the shortlist is formed, and it is formed mostly from reviews and independent coverage, not from what the brand says about itself.(Chen et al., University of Toronto, 2026, pp. 3-4)

Transactional: owned pages take over

For purchase questions (“buy … near me”), brand sites become the largest source at 52.9 %. Every AI system raised its share of brand citations sharply, to between 52 % and 68 %. At this point the answer depends on whether the brand’s own pages are findable and current: prices, availability, where to buy.(Chen et al., University of Toronto, 2026, pp. 3-4)

generative engines systematically privilege earned and brand-owned content while under-representing social and community perspectives, and their source composition varies far more sharply across intents than Google’s relatively stable profile.
Chen et al., University of Toronto, 2026, p. 4

4The engines differ, and Google most of all

Across all three intents, the profiles of the individual systems are far apart. Google is the most balanced and the only one in which social platforms carry real weight. Among the AI systems, Claude 4.5 Sonnet and GPT-4o lean hardest on earned media; Gemini 2.5 Flash, which answers with Google Search grounding, gives brand sites as much weight as earned ones.(Chen et al., University of Toronto, 2026, pp. 2-4)

Earned and brand sources per system, all intents

Share of cited sources in %, column “All Intents” of fig. 2; 300 consumer-electronics questions

  • Earned
  • Brand (owned)
Google Search
Earned: 40.7 %, Brand (owned): 25.7 %
GPT-4o
Earned: 56.9 %, Brand (owned): 35.6 %
Claude 4.5 Sonnet
Earned: 64.5 %, Brand (owned): 34.4 %
Perplexity Sonar Pro
Earned: 49.9 %, Brand (owned): 38.7 %
Gemini 2.5 Flash
Earned: 46.2 %, Brand (owned): 46.4 %

The remainder of each row is social: 33.6 % for Google, against 1.1 % (Claude) to 11.4 % (Perplexity) for the AI systems.

Source: Chen et al., University of Toronto, 2026, p. 4, fig. 2

Two cautions before reading too much into single engines. First, the authors report that Claude 4.5 Sonnet returned no links for most informational and transactional questions until it was explicitly prompted to search, although it was queried in web-enabled mode; its shares rest on that behaviour. Second, these are one configuration of each engine at one point in time. The practical conclusion is not “optimise for engine X”, but measure each engine you care about separately, because a brand can be well cited in one and absent in another.(Chen et al., University of Toronto, 2026, p. 3)

A second finding of the study explains why earned coverage matters even more for smaller brands. In a case study with GPT-4o alone, rankings of well-known brands barely changed when the retrieved evidence was shuffled or altered; on average 16 % of the entities in the generated rankings did not appear in any retrieved snippet, so the model filled in from what it already knew. For niche entities the retrieved pages drove the answer directly. A brand the model does not already know depends on what it can find about it at the moment of the question.(Chen et al., University of Toronto, 2026, pp. 8-9)

current AI search behavior indicates that specific source types, particularly earned and owned media, contribute more strongly to search presence than others.
Chen et al., University of Toronto, 2026, p. 9

5What the study does not show

The paper measures what engines cite. It does not test which tactics make a page more likely to be cited.
Everything-PR, Everything-PR, June 2026, What are the limits of the study?

6What communication teams should pay attention to

  1. 1

    Map your questions by intent

    List the questions people ask an assistant about your category and sort them into learning, comparing and buying. Each group is answered from a different mix of sources, so each needs its own measurement and its own owner.

    Chen et al., University of Toronto, 2026, p. 4, fig. 2

  2. 2

    Win the comparison stage with earned media

    Comparison questions are answered mostly from independent media and reviews. Find out which outlets the engines cite for your category, and which journalists there cover it, and direct press work to them.

    Chen et al., University of Toronto, 2026, pp. 3-4

  3. 3

    Keep owned pages ready for the purchase question

    At the buying stage brand sites lead. Product, price, availability and where-to-buy pages have to be current and clear enough to be quoted.

    Chen et al., University of Toronto, 2026, pp. 3-4

  4. 4

    Keep independent coverage current

    Everything-PR reads the study’s freshness and source-type results together: earned media and fresh content were cited more often by the AI engines than by Google, which in its view "supports keeping independent coverage current".

    Everything-PR, June 2026, What does this mean for communications teams?

  5. 5

    Tell paid from earned in your own reporting

    Sponsored pieces on cited outlets will sit in the same bucket as editorial coverage in any site-level count. Record which is which, so you can see whether paid placements ever reach an AI answer instead of assuming they do (why the study cannot tell).

  6. 6

    Do not treat social as a GEO lever by default

    Social platforms carry a third of Google’s sources in the study but between about 1 % and 11 % of the AI systems’. Social still matters for reach and reputation; it is a weak route into AI answers on this evidence.

    Chen et al., University of Toronto, 2026, p. 4, fig. 2

  7. 7

    Measure per engine, per market, over time

    Overlap with Google is low and engines differ widely, so a Google ranking says little about AI visibility. Track whether each engine names and cites the brand, in each market you sell in, and repeat it, because the snapshot ages.

    Chen et al., University of Toronto, 2026, pp. 2-4

7How Acontema helps

Acontema brings the two sides of this picture together: what AI answers say and cite, and the media coverage they draw on.

Know which questions matter and whether the brand appears in the answers
AI visibility (GEO) with Prompt Research to find the questions and Prompt Tracking to measure whether AI answers name and cite the brand, per market.
Earned media carries the comparison stage
Media monitoring of earned and paid coverage, with a journalists catalogue matched by what each journalist has covered.
Paid and earned must not be mixed up
Monitoring distinguishes paid from earned coverage, with syndicated and paid placements detected, so the share that was actually earned stays visible (see the note).
Social sits in both owned and earned
The brand’s own social channels and the comments under its posts are collected and analysed separately from what others publish.
Results change over time
Monthly reports bring AI visibility and coverage together, so changes show up as a trend rather than a one-off audit.

Notes

Owned, earned and paid: how the study’s categories map

Communication teams usually plan in owned, earned and paid media. The study’s three types do not line up one to one with that scheme. The classification is by the site a source sits on, not by how the text came about, so the mapping below is our reading of the study’s definitions, not a claim the authors make.

Brand: official company-owned domains
Owned. The brand’s website, product pages, support and shop pages.
Earned: independent media and review outlets
Earned, and in practice also some Paid. Editorial coverage and reviews are earned. A sponsored article or advertorial on a publisher’s domain sits on the same independent site, so a site-level label would most likely count it here too. The study does not separate paid from earned content and reports no figure for paid placements.
Social: community or user-generated platforms
Split. The brand’s own accounts on these platforms are owned channels; posts, threads and reviews by other people are earned mentions. The study counts both as Social.
Not covered
Paid search and display ads do not appear as a category. The study measures the sources an answer cites, not advertising shown around it.

Sources

Quotes are verbatim. Page numbers refer to the printed page of the PDF. Statements about the platform describe what it does today.

See it on your own brand.

Start free trial

14-day free trial. No credit card required.