A multi-engine, multi-market benchmark study of how six AI search engines — ChatGPT, Perplexity, Gemini, Microsoft Copilot, Google AI Overview and Google AI Mode — answer buyer-intent questions about enterprise bot management, bot mitigation and automated-traffic defence across the United States, Canada, India and Germany.
A multi-engine benchmark of how six AI search engines answer buyer-intent questions about enterprise bot management. Covers 240,000 AI engine responses drawn from 10,000 queries across ChatGPT, Perplexity, Gemini, Microsoft Copilot, Google AI Overview and Google AI Mode in four markets, together with 3,906,000 cited sources classified by source type and 84 distinct product entries. Reports vendor share of voice, first-mention position, cross-engine cited-source overlap measured by Jaccard similarity, market-by-market variance, and the share of cited pages that sit on vendors' own sites.
The six engines converge on the same four-vendor core while drawing on largely separate evidence bases: cited-domain overlap averages a Jaccard coefficient of 0.17, so a visibility programme measured on one engine does not generalise to another. The report quantifies that gap, separates mention volume from ordinal position, and finds 47.3% of cited pages on vendors' own sites with only 2.0% dead.
The opening pages of the PDF, reproduced here: study scale, abstract, table of contents and introduction. Download the full report for the results, discussion and appendices.
Responses recorded and sources cited, per engine.
| AI engine | Responses | Cited sources | Markets |
|---|---|---|---|
| ChatGPT | 40,000 | 1,950,000 | 4 |
| Microsoft Copilot | 40,000 | 155,000 | 4 |
| Gemini | 40,000 | 166,000 | 4 |
| Google AI Mode | 40,000 | 766,000 | 4 |
| Google AI Overview | 40,000 | 469,000 | 4 |
| Perplexity | 40,000 | 400,000 | 4 |
| Total | 240,000 | 3,906,000 | 4 |
Security teams that buy automated-traffic defence now begin the evaluation inside a generated response, so the vendor set an AI engine returns forms the consideration set before any vendor is contacted. This study measures how six commercial AI search engines answer buyer-intent questions about enterprise bot management. Each engine received 10,000 enterprise buyer-intent queries in each of four markets during a single collection window in August 2026, producing 240,000 responses that contained 3,906,000 cited sources and covered 84 distinct product entries. For every response the study recorded length, cited pages and their source type, page reachability, hedging, recency anchoring, truncation, refusal, and each brand mention with its ordinal position. The engines converge on a core: 10 of 37 brand strings were named by all six engines, and the four most-mentioned — Cloudflare, Akamai, DataDome and Imperva — took 55.8% of all mentions, with 18 strings named by a single engine. Cloudflare was named most often, at 232,000 mentions, and opened the response in 5 of the six engines; DataDome opened ChatGPT, where it averaged position 1.5 against Cloudflare's 3.1. The engines cite largely separate domains, with a mean pairwise overlap of 0.17, and no two engines share a most-cited domain. 47.3% of everything they cite sits on a vendor's own site, but vendor sites are the largest class for only 1 of the six engines. Dead links are rare, at 2.0% of cited pages. Cloudflare led every market; the order below it changed by market within the same four brands.
A security team evaluating bot management now begins inside a generated response. The buyer asks an AI search engine which products separate automated traffic from human traffic at the edge, protect login and checkout flows from credential stuffing and account takeover, and admit the automated agents the business wants while refusing the ones it does not. The engine returns a short ordered list of named vendors with a paragraph of justification for each. That list is not a ranking of the market. It is the output of one system reading one collection of web sources at one moment, and it arrives before the buyer has opened a single vendor site.
The structural consequence is narrow. A vendor absent from the returned list is not ranked low; it is not present. A generated response has no second page, so the buyer has no mechanism equivalent to scrolling further. Where a search results page offered ten positions and a set of related queries, a generated response typically names between three and six vendors and closes. Visibility in this setting is closer to inclusion than to ranking, and the difference between being inside and outside the returned set is discontinuous.
The measurement problem that follows is that no single engine describes the market. Engines differ in the sources they read, the length at which they answer, and the order in which they name vendors. A vendor that measures its position in one engine learns about that engine. This study measures six at once, in four markets, over one collection window, so that agreement and divergence between them can be separated from the behaviour of any one system.
Enterprise bot management is a suitable instrument for this measurement. The vendor set is mature and well documented, so the engines have abundant published material to draw on, and 10 of the 37 brand strings recorded were named by every engine. A category with a settled core isolates engine behaviour from genuine market churn.
The category also has a stable and publicly articulated capability vocabulary. Across the corpus the engines returned the same selection factors: behavioural and machine-learning detection, device fingerprinting, low false-positive rates, real-time scoring, protection for login and checkout flows against credential stuffing and account takeover, integration with the content-delivery network or web application firewall already in place, and selective admission of verified automated agents. These criteria are specific enough that responses can be compared to each other rather than to a general description of security software.
The category also splits cleanly along one design axis the engines return repeatedly: platforms native to a content-delivery network against independent layers that sit in front of any network. That split gives the engines a genuine choice to make when a query asks for a single recommendation, and §4.9 records that they make it differently.
This study provides four things a single-engine measurement does not. It establishes how far six engines agree on the vendor set in one category, separating the shared core from the long tail each engine holds alone. It measures the overlap between the engines' evidence bases directly, as a pairwise coefficient over cited domains, and classifies every cited page by source type, so that agreement on vendors can be set against what the engines actually read. It reports mention volume, first-mention position and mean ordinal position as three separate quantities, which allows them to be shown moving independently. And it tests the vendor set across four markets in the same window, which distinguishes market-specific results from a set the engines return everywhere. All four rest on a single corpus of 240,000 responses collected under one design.
Research questions, methodology, results, discussion and appendices.
Start your 7-day free trial and get your AI Visibility Score in about a minute. See exactly where you stand, where competitors are beating you, and the ranked fixes to get into the answer. Cancel anytime before the trial ends.
7-day free trial. Cancel anytime before it ends. Trusted by 500+ B2B SaaS teams.
