Who AI Recommends for GEO Agency Searches

Key Findings

We collected 12,000 answers from ChatGPT, Google AI Mode, Gemini, Perplexity and Microsoft Copilot across 80 prompts, three markets and two waves. Recommendation rankings use 10,500 answers after excluding the 1,500 brand-validation answers in group G5; those answers were used only for factual-accuracy checks.

1. 01AI is willing to name names. In the ranking-eligible set, 74% of answers recommended at least one specific agency (95% CI 72%-76%; n = 10,500), with an average of 2.3 agencies per non-refused answer.
2. 02A small group takes most of the visibility. We recorded 142 distinct agency names across the full collection, while the pooled top 10 captured 48% of recommendations in the 10,500-answer ranking-eligible set (95% CI 45%-51%).
3. 03The engines rarely agree. Average overlap between any two engines' top-10 lists was 0.41 (95% CI 0.37-0.45), less than half on the Jaccard scale.
4. 04Third-party lists drive recommendations. In the ranking-eligible set, listicles made up 40% of cited sources (95% CI 37%-43%), versus 15% from agencies' own websites.
5. 05Ask twice, get a different answer. Across the ranking-eligible prompts, five runs of the same prompt produced an average overlap of 0.52 (95% CI 0.49-0.55). Sixty percent of wave-one top-10 agencies remained three weeks later.
6. 06Listicle presence beats rankings. Recommended agencies appeared in a median of 8 listicles, compared with 2 for Google-ranking agencies that were rarely recommended.
7. 07Ranking on Google is not enough. Thirty-nine percent of agencies ranking in Google's top 10 were recommended in less than 1% of eligible AI answers.
8. 08AI sometimes invents agencies. Nine of 142 distinct names, or 6.3% (95% CI 3.0%-11.7%), could not be verified as real companies.

Why We Ran This Study

Buyers increasingly ask an AI assistant before they ask Google. When the question is "which agency should I hire?", the answer is no longer a page of ten blue links. It is a short list of names, written by a model, often with no click required.

The click is shrinking fast. Ahrefs found that the top-ranking page earns about 58% fewer clicks on keywords that show an AI Overview, comparing December 2023 with December 2025 across 300,000 keywords (Ahrefs). Pew Research Center found that users clicked a traditional result in 8% of visits when an AI summary appeared, versus 15% without one (Pew Research Center).

Being named inside the answer is what counts now. Seer Interactive reported that brands cited in an AI Overview earned 35% higher organic click-through than those that weren't (Seer Interactive). Yet our earlier test of 216 brands found that 62% of AI answers recommending a brand cited none of that brand's own pages (Freeways). Recommendations are built mostly from what others say about you.

That raises an obvious question for anyone hiring a GEO agency: if these agencies sell AI visibility, which of them does AI actually recommend, and why? We set out to measure it.

Methodology at a Glance

We collected 12,000 AI answers between 14-18 September 2026 and 5-9 October 2026, then audited the public footprint of 100 agencies.

Part 1: How AI Search Changed Buyer Shortlists

AI answers now sit between the buyer and the vendor's website. For a service purchase like hiring an agency, that means the shortlist is often formed before the buyer visits a single site.

AI search is now mainstream. Google reported more than 2 billion monthly AI Overviews users in July 2025 (Google Q2 2025) and more than 1 billion monthly AI Mode users in Q2 2026 (Google Q2 2026). OpenAI reported more than 1 billion weekly active ChatGPT users (OpenAI). McKinsey found that half of surveyed consumers intentionally use AI-powered search (McKinsey).

Fewer clicks reach the ranking pages. Beyond the Ahrefs figure above, Seer Interactive's tracking shows organic click-through on queries with an AI Overview declining sharply, while brands cited inside the overview hold up better (Seer Interactive). Different studies report different sizes of decline because they compare different periods, query types and positions; the direction is consistent.

The AI answer works like a shortlist. In Pew's data, users clicked a source inside an AI summary in only about 1% of visits (Pew Research Center). For most buyers, the names in the answer are the names that get considered.

Measurement practices are still catching up. The practical gap is not whether buyers use AI search, but whether brands track how often they are mentioned, cited and described accurately across engines.

For GEO agencies, this is a credibility test: they sell AI visibility, so their own visibility is public evidence.

Part 2: GEO Agencies Recommended by AI

Across the full 12,000-answer collection, AI named 142 distinct agencies. For rankings, we excluded the 1,500 G5 brand-validation answers and analyzed 10,500 ranking-eligible answers. The top 10 agencies' individual mention rates sum to 110.4%; divided by the observed 2.3 agencies per non-refused answer, that equals 48.0% of recommendation mentions.

The most-recommended GEO agencies

Each engine builds a different shortlist

The pooled top-10 share (48%) uses one common cross-engine list and is calculated from the T1 mention-rate total (110.4% / 2.3 agencies per non-refused answer). The percentages in Table T2 use each engine's own top 10, so the pooled and engine-specific shares answer different questions and are not directly additive.

The average top-10 overlap between two engines was 0.41. The closest pair was ChatGPT and Perplexity (0.52); the most different was Google AI Mode and Copilot (0.28). In practice, no engine's top 10 reliably predicted another engine's shortlist.

Markets see different agencies

Only 30% of agencies in the top 10 appeared in all three markets' top 10. Local market context therefore changed most shortlists, even when prompts were translated only by market name.

Specific questions surface specialist agencies

Constrained prompts produced an average of 4.2 distinct agency names per prompt, compared with 4.9 for generic best-agency prompts. However, 61% of the agencies surfaced by constrained prompts did not appear in the generic set, consistent with the hypothesis that niche positioning can surface smaller specialists even when each prompt returns fewer names.

Even when the question didn't ask for a recommendation, such as "What does a GEO agency do?", 19% of answers named at least one agency. These unprompted mentions are the strongest signal that an engine treats an agency as an authority on the topic.

Part 3: Sources Behind AI Agency Recommendations

When AI recommended a GEO agency, it mostly relied on what third parties wrote about that agency, not on the agency's own website.

What the wider research says

Large citation studies point the same way for brands in general:

• Each engine has its own favorite sources. Profound's analysis of 680 million citations from August 2024 to June 2025 (Profound) found that Wikipedia accounted for 7.8% of all ChatGPT citations, while Reddit led Google AI Overviews at 2.2% and Perplexity at 6.6%. The much larger figures often quoted for these domains refer only to share within each platform's ten most-cited sources, not all citations.
• Earned media dominates. Muck Rack's analysis of more than 1 million AI-cited links (Muck Rack) found that 89% came from earned media and 27% from journalistic content.
• Retrieval is selective. Ahrefs' study of 1.4 million prompts (Ahrefs) found that relevance to ChatGPT's fan-out queries distinguished cited pages from retrieved but uncited pages; appearing in search retrieval did not guarantee citation.
• Being talked about matters. Ahrefs found branded web mentions correlated at 0.664 with AI Overview visibility (Ahrefs).

What we found for GEO agencies

Perplexity leaned on listicles the most, at 48% of its citations. Agency-owned websites accounted for 15% across engines. Among the ten most-cited domains were neutral directories and community platforms such as Clutch, GoodFirms, DesignRush, LinkedIn and Reddit. Four of the ten were agency-owned publishers of rankings that included their own firms, so source ownership was audited separately from citation volume.

Part 4: Stability of AI Recommendations

AI recommendations are not a fixed ranking. On the 10,500 ranking-eligible answers, the same prompt asked five times in fresh sessions returned an average overlap of 0.52 between runs, where 1 means an identical list.

This matches what we saw in our earlier test of 216 brands, where only 34% of unique cited URLs appeared in all three runs of the same prompt on the same platform (Freeways).

Google AI Mode was the most consistent and Gemini the least. The variation between runs shows that a single run is too unstable to treat as a ranking.

Two practical consequences follow:

• A single screenshot proves very little. Any agency report built on one run of one prompt can be off by a wide margin. In our data, you needed about 4 runs per prompt for the top-5 list to stabilize.
• Visibility should be measured as a rate, not a rank. For example, First Page Sage appeared in about 17 of every 100 ranking-eligible answers in this analysis. That is more informative than saying an agency was ranked number one in a single ChatGPT response.

Part 5: Signals Shared by Recommended Agencies

To understand why some agencies get recommended, we compared 60 frequently recommended agencies with 40 agencies that offer GEO services and rank on Google, but were recommended in less than 1% of answers.

The strongest signal was listicle inclusion count (rho = 0.70). Being present in independently published lists had a stronger association with G5-excluded recommendation rate than any on-site signal.

Weak or no relationship: public pricing and llms.txt adoption. Neither signal had a statistically distinguishable positive association with recommendation rate.

Google rank was not interpreted as a clean pooled predictor because the control group was selected partly for ranking well on Google. Engine-specific associations were directionally stronger for Google AI Mode, but they remain descriptive and are vulnerable to this selection effect.

Ranking alone is not enough. Of the agencies ranking in Google's top 10 for 'generative engine optimization agency' in at least one market, 39% were recommended in less than 1% of eligible AI answers. Because ranking was part of control selection, this comparison should not be read as a causal estimate.

In a combined model, listicle count, referring domains, review volume, original research, and recent news mentions together explained 45% of the variation in the G5-excluded mention rate. These are correlations, not proof of cause: agencies that are widely listed may also be larger, older, or simply better known.

How this fits with content-level research. The original GEO study from Princeton, Georgia Tech, IIT Delhi and the Allen Institute found that adding statistics, citations and expert quotations can improve a page's visibility in generative answers. The agency-level analysis suggests that page-level optimization works best when third-party sources already establish the brand as a relevant entity.

Part 6: Errors in AI Agency Recommendations

AI recommendations are not always accurate. Of the 142 distinct agencies AI recommended, 6.3% could not be verified as real companies, and a further 4.9% were real but don't offer GEO services, such as software tools or agencies focused only on paid ads.

Even for the most-recommended agencies, details were often off. We checked every claim AI made about the top five agencies in each market against their official websites: 13% of those claims were incorrect.

This is consistent with our earlier test of 216 brands across 12 industries, where 18% of AI answers mentioning a brand contained at least one factual error, such as an outdated price or a wrong service area (Freeways).

For buyers: treat an AI shortlist as a starting point and verify every agency on its own site and on independent review platforms. For agencies: inaccurate AI descriptions are a fixable problem, usually by making facts consistent across your website, directory profiles and the third-party pages AI cites.

Implications for Brands and Buyers

If you want AI to recommend your brand

The agencies AI recommends are not simply the ones with the best websites. They are the ones other sources talk about. The same pattern holds well beyond GEO agencies.

1. 01Get onto the lists AI reads. Find the listicles, directories and comparison pages AI cites for your category, and earn a place on them with real proof: reviews, case studies, data.
2. 02Build a footprint across engines. Each engine leans on different sources, so a single-channel strategy will leave gaps. Check where you appear on each engine, not just ChatGPT.
3. 03Keep your facts consistent. Inaccurate AI descriptions usually trace back to conflicting information across your site, profiles and third-party pages.
4. 04Measure rates, not ranks. Run each important prompt several times, on several engines, and track the share of answers that recommend you.
5. 05Give AI something to cite. Original data, clear answer blocks and named expertise make your pages easier to quote when you are retrieved.

If you are hiring a GEO agency

An agency that sells AI visibility should be able to show its own. Ask these questions before signing:

• How often do AI engines recommend you, and how did you measure it? Look for multiple engines, multiple runs and a defined prompt set, not a single screenshot.
• Can you show before-and-after AI visibility data for a client? Case studies should report citation or mention rates, not just rankings or traffic.
• Do you work on third-party presence, or only on our website? Our data suggests on-site work alone is unlikely to be enough.
• What will you report every month? Expect mention rate, citation share and accuracy by engine.
• Do you guarantee a position in ChatGPT? Given how much answers vary between runs, a guarantee is a red flag.

Pricing varies widely. Published market guidance ranges from roughly $1,500 to more than $50,000 per month, depending on scope. In the agency audit, 18 agencies published usable prices and the median starting price in the analysis was $4,500 per month.

FAQ

GEO Agency Most Often Recommended by ChatGPT

In our study, ChatGPT most often recommended First Page Sage, appearing in 16.5% of ranking-eligible ChatGPT answers. Results varied by market and between runs, so treat any single AI answer as a sample, not a verdict.

Agreement Across AI Engines

No. The average overlap between two engines' top-10 lists was 0.41 on a scale where 1 means identical. Each engine relies on a different mix of sources.

How AI Selects Agencies

AI engines retrieve web pages and summarize what they find. In our data, the agencies recommended most often were those that appear on many third-party lists and review platforms; listicle inclusion count had the strongest relationship with recommendations.

Whether Google Ranking Is Enough

Not on its own. Thirty-nine percent of agencies ranking in Google's top 10 for "generative engine optimization agency" were recommended in less than 1% of eligible AI answers.

Unverifiable Agency Recommendations

Yes, occasionally. 6.3% of the distinct agency names AI produced could not be verified as real companies. A further 4.9% referred to real organizations that did not offer GEO services. Always verify an agency independently.

Measuring Brand Visibility in AI

Use a fixed set of prompts that reflect real buyer questions, run each several times on each major engine, and track the share of answers that mention or recommend you. Single screenshots are unreliable because answers change between runs.

Cost of a Generative Engine Optimization Agency

Estimates vary widely by scope. WebFX puts agency GEO services at roughly $1,500 to more than $50,000 per month.

Methodology, Limitations and Disclosure

Prompt design

We wrote 80 prompts in English, grouped by buyer intent: education (10), generic recommendation (15), constrained recommendation (35, covering industry, budget, company stage, service focus and platform), evaluation and comparison (10), and brand validation (10). Prompts mentioning a market used 'Singapore', 'Australia' or 'the United States'. Brand-validation prompts used the five most-recommended agencies per market from a pilot run, fixed before wave one. Group G5 was used only to test factual accuracy and was excluded before calculating mention rate, top-3 rate, share of voice, source mix (T3), stability (T5), signal relationships (T6) and all ranking tables. The full prompt list is included in the research package.

Data collection

• Engines: ChatGPT with search, Google AI Mode, Gemini, Perplexity and Microsoft Copilot. Google AI Overviews were measured separately using 20 keyword-style queries per market and wave; an AI Overview appeared for 58% of them.
• Sessions: every run in a fresh session, single turn, with location set to the target market; logged out, or using research accounts with memory and personalization turned off.
• Repetition: 5 runs per prompt, engine, market and wave; prompt order randomized within each run.
• Waves: wave one 14-18 September 2026, wave two 5-9 October 2026.
• Collection mode: 62% monitoring tool, 28% API with web search, and 10% manual UI. On a calibration sample of 8 prompts run through both the API and the consumer interface, agency lists overlapped at 0.82.

Coding

Every agency named in an answer was coded as recommended, mentioned or negative, with its position and sentiment. Every cited URL was classified into one of nine source types. Two coders independently coded 10% of answers; agreement on key fields reached Cohen's kappa = 0.86. Agency names were normalized using a shared dictionary, and agencies were verified against their websites and company profiles.

Signal audit

We audited 100 agencies during a single week (28 September-2 October 2026): the most-recommended agencies from wave one, plus a control group of agencies offering GEO services that rank in Google's top 30 for GEO agency keywords but were recommended in less than 1% of answers. All signals came from public sources, recorded with evidence URLs.

Analysis

Confidence intervals were calculated by bootstrapping at the prompt level, since repeated runs of the same prompt are not independent. Unless a section explicitly concerns factual accuracy, every reported recommendation rate, ranking, citation-source share, stability estimate and signal relationship uses the same 10,500 G1-G4 ranking-eligible answers; the 1,500 G5 brand-validation answers are excluded. Relationships between signals and mention rate are reported as Spearman correlations and a multivariable model with a limited number of signals. We re-ran key results excluding Freeways and using only interface-collected data; the ordering and effect sizes did not change materially, and no headline estimate moved by more than 2 percentage points.

Limitations

• AI answers change with model versions, time, location and personalization. This analysis uses a September-October 2026 collection window.
• API responses can differ from what users see in the interface; our calibration sample estimates that gap.
• We tested English prompts in three markets; results may differ in other languages and regions.
• Candidate discovery combined neutral directories, engine-generated longlists and agency-authored rankings. Because several agencies publish lists that include themselves, the candidate pool may still overrepresent firms with strong listicle distribution.
• Signal results show correlation, not causation, and the audit sample is limited to roughly 100 agencies.

Disclosure

Freeways is a GEO agency. We report Freeways' result in this analysis alongside every other agency and rerun the analysis without it as a sensitivity check. We do not name agencies in sections on errors or inaccurate descriptions.

Sources

• Ahrefs: AI Overviews Reduce Clicks by 58%
• Fast Company Featured AI Search Study from Seer Interactive
• Google: Alphabet Q2 2025 CEO remarks on AI Overviews adoption
• Google: Alphabet Q2 2026 CEO remarks on AI Mode adoption
• OpenAI: More than 1 billion weekly active ChatGPT users
• McKinsey: New front door to the internet - winning in the age of AI search
• Pew Research Center: Google users click less when an AI summary appears
• Muck Rack: What Is AI Reading?
• Profound: AI Platform Citation Patterns
• WebFX: Generative Engine Optimization Cost
• Aggarwal et al.: GEO - Generative Engine Optimization
• Ahrefs: Why ChatGPT Cites One Page Over Another - study of 1.4M prompts
• Ahrefs: AI Overview Brand Visibility Factors (75K brands)

Candidate Discovery Sources

• First Page Sage 2026 GEO agency ranking
• LoudFace 2026 GEO and AEO agency review
• WebFX 2026 GEO company review
• Clutch GEO provider category
• RankTracker 2026 GEO and AEO agency review
• Canlah AI global market review
• DesignRush GEO agency directory
• GoodFirms GEO agency directory

Engine-generated candidate discovery: neutral category prompts were run without agency names; every returned organization was then verified against its official site before inclusion.

Appendix A: Top 30 Agencies by Measured Mention Rate

Purpose. This table reports the 30 agencies with the highest measured mention rates in the 10,500 ranking-eligible answers. It is a results table, not the 100-agency signal-audit comparison group and not an independent quality ranking or endorsement. Group G5 brand-validation prompts are excluded; rates were recalculated using the reduced denominator rather than copied from the 12,000-answer collection.