Which Attack Surface Management Platforms AI Engines Recommend: A 240,000-Response Study

When an enterprise security buyer opens an AI engine before a vendor site, the shortlist the engine returns shapes the consideration set before anyone fills in a contact form. We wanted to know what those engines actually say in one category, measured at scale rather than inferred from a handful of prompts.

Key Takeaways

• CyCognito recorded 160,000 mentions and Palo Alto Networks 157,000, the two highest of 87 brand strings. Seven brands were named by all six engines; 47 were named by only one.
• Mention volume and position move independently. CyCognito opened the response in three engines, but Palo Alto Networks was read earlier on average in four. In ChatGPT its mean ordinal position was 1.4 against CyCognito's 4.5, where 1 means named first.
• Cited-domain overlap between engines averaged 0.16 as a Jaccard coefficient, where 0 is disjoint and 1 is identical. No pair from different vendor families exceeded 0.20, and no two engines shared a most-cited domain.
• 51.3% of all cited pages were on a vendor's own site. Microsoft Copilot was the exception at 23.1%.
• Only 0.4% of cited pages were dead.
• The same two brands led all four markets, with the order reversing only in Germany.

What We Analyzed and How

We issued 10,000 enterprise buyer-intent queries about attack surface management platforms to six AI engines: ChatGPT, Perplexity, Gemini, Microsoft Copilot, Google AI Overview, and Google AI Mode. Every query ran in four markets (United States, Canada, India, Germany), producing 40,000 responses per engine and 240,000 in total during August 2026.

For every response we recorded length, cited sources and the type of site each one sits on, whether each cited URL still resolved, each brand named, and the ordinal position of that brand's first appearance. The corpus contained 1,172,000 cited sources and 124 distinct product entries once plan, edition and category-noun spellings of one product were merged.

Two limits are worth stating up front. The engines are commercial services without pinned model versions, so a repeat run would not reproduce identical responses. And the query set is purposive rather than randomly sampled, so we report descriptive statistics only and make no claims of statistical significance. Full methodology and validity limits are in the complete benchmark report.

Which Vendors Do AI Engines Name Most in Attack Surface Management?

CyCognito led the category with 160,000 mentions, followed by Palo Alto Networks at 157,000 and Microsoft at 115,000. Censys was 2,000 behind Microsoft at 113,000. Those four brands took 43.8% of all mentions across the 87 brand strings recorded. Seven brands were named by all six engines: CyCognito, Palo Alto Networks, Microsoft, Tenable, CrowdStrike, IONIX and Axonius.

The core is small and the tail is long. 47 of the 87 brand strings were named by exactly one engine, none of them above 9,000 mentions, and several of the strings are vulnerability management, threat intelligence or ticketing vendors that an engine pulled into the category on its own. Between the two sits a band of brands whose coverage is uneven in ways a single number hides. Bitsight recorded 19,000 mentions in Gemini and 1,000 in Google AI Overview, and did not appear in Microsoft Copilot at all. Censys, Rapid7, Bitsight, Qualys, Wiz and runZero were each recorded in five engines rather than six.

Our finding: across 240,000 analyzed responses, three engines opened with CyCognito, two opened with Palo Alto Networks, and one opened with Microsoft. That split follows mention volume. Mean ordinal position does not: Palo Alto Networks was read earlier than CyCognito in four of the six engines.

First Mention and Mean Position Tell Different Stories

The brand mentioned most is not reliably the brand named first. CyCognito held the first-mention position in three engines, all three operated by Google: Gemini, Google AI Mode and Google AI Overview. Palo Alto Networks held it in ChatGPT and Microsoft Copilot. Microsoft held it in Perplexity.

Mean ordinal position tells a third story. Palo Alto Networks was read earlier on average than CyCognito in four of the six engines, including Google AI Mode, where CyCognito opens the response most often but Palo Alto Networks averages 1.8 against 2.2. The widest gap was in ChatGPT: Palo Alto Networks averaged 1.4, the earliest mean position among the three leading brands in any engine, while CyCognito averaged 4.5, the latest.

Microsoft Copilot shows the same split in plain mention counts. CyCognito recorded 32,000 mentions there against 29,000 for Palo Alto Networks, and Palo Alto Networks still took the opening slot. At product level the pattern repeats: Cortex Xpanse recorded 149,000 mentions at a mean rank of 1.8, the earliest of the ten most-mentioned products, while CyCognito's own product entry recorded 160,000 at 2.6.

These are three different quantities. Mention volume records how often a vendor enters the response. First-mention position records which vendor opens it. Mean position records where a vendor lands once named. If your visibility report gives you one number per brand, it has already chosen which of the three to hide. Our guide to how to measure AI share of voice across engines sets out how to keep them separate.

Which Sources Do AI Engines Cite for ASM Recommendations?

ChatGPT supplied 357,000 citations and Google AI Overview 351,000, together 60.4% of the 1,172,000 cited sources in the corpus. Microsoft Copilot supplied 156,000, Google AI Mode 175,000, Gemini 112,000 and Perplexity 21,000.

More than half of that evidence is the vendors describing themselves. 51.3% of all cited pages sit on a domain belonging to a vendor named in the response. Analyst and review sites carry another 11.1%. The residual class, pages outside every type we track, holds 34.5%.

The engines are nothing alike on this measure. Perplexity drew 81.0% of its citations from vendor sites and ChatGPT 69.7%. Gemini followed at 56.2%, Google AI Overview at 49.0% and Google AI Mode at 36.6%. Microsoft Copilot is the opposite case: 23.1% from vendor sites and 71.8% from the residual class, led by two small third-party security sites at 24,000 and 17,000 citations that appear in no other engine's top three. Google AI Mode leaned hardest on analyst and review pages, at 20.0% of its citations, and its single most-cited domain was an analyst firm.

The most-cited domains confirm the separation. ChatGPT's were Palo Alto Networks' own domain at 59,000 citations, an analyst firm at 42,000, and Tenable's domain at 30,000. Google AI Overview's were CyCognito's domain at 53,000, IONIX's at 30,000 and Palo Alto Networks' at 27,000. No two engines shared a most-cited domain. How each engine settles on its sources is a question we have looked at separately in how the major AI engines decide which sources to cite.

Measured as a Jaccard coefficient over cited domains, where 1.00 means identical sets and 0.00 means nothing shared, overlap between engines averaged 0.16 across the 15 pairs. The three highest values were the three pairs of Google-operated engines, at 0.28, 0.27 and 0.38. No pair from different vendor families exceeded 0.20, and Microsoft Copilot recorded the lowest coefficient with every other engine, from 0.04 to 0.11.

This is the clearest case in the data for measuring more than one engine. Six engines returned substantially the same core vendors while reading substantially different sets of domains, so a source list that serves ChatGPT describes little of what Microsoft Copilot or the Google surfaces read.

How Reliable Are the Sources AI Engines Cite?

Of the 1,172,000 cited URLs we tested, 4,240 did not resolve, a dead-link rate of 0.4%. No engine exceeded 0.6%, the rate for ChatGPT, and three engines recorded no dead links at all: Microsoft Copilot, Gemini and Perplexity. Google AI Overview stood at 0.3% and Google AI Mode at 0.6%. ChatGPT alone accounted for 2,140 of the 4,240 dead pages, half the total, and it also cites the most.

The definition matters. Many sites answer an automated request with a 403 because the page sits behind a bot wall. The page exists; the crawler was turned away. Count those as dead and the rate balloons. We count a page as dead only when it returns not-found, gone or a server error, or its host cannot be reached after a retry. On that definition, the pages these engines lean on are almost all there, which fits the previous finding: most of them are the vendors' own pages, and vendors keep their own sites up.

Source age is not reported, because fewer than a fifth of the cited pages declared a publication date. For a vendor, the practical reading is that link decay is not the lever in this category. The inclusion problem is.

Do AI Recommendations Change by Country?

Not in membership. The same two brands led all four markets, and no market returned a vendor absent from the other three in its leading positions.

Brands ranked by mentions within each market; n = 60,000 responses per market. Source: GrackerAI AI Search Visibility Benchmark Series, August 2026.

CyCognito was first in the United States, Canada and India with Palo Alto Networks second. Germany reversed the pair. The third position alternated between Censys and Microsoft. Below the leading group the stability held: 29 of the 124 product entries were recorded in all four markets, including every one of the ten most-mentioned.

That is tighter than our analysis of what AI engines recommend in network security, which found recommendations diverging by region, though it covered seven engines across ten markets, so the difference may be scope rather than category. The budget implication here is direct: a vendor missing from the shortlist in one of these four markets is missing in all of them, and a localized landing page is not what fixes it.

Engine choice mattered far more than geography. Google AI Overview averaged 232.3 words per response and ChatGPT 647.4, a factor of 2.8. Two buyers asking the same question of different engines receive answers built to different shapes from different evidence.

Where Did the Engines Disagree?

The first disagreement was over asset discovery after a merger or acquisition. ChatGPT returned Cortex Xpanse as the product for that use case. Gemini and the two Google search surfaces more often led with CyCognito, citing its seedless attribution of subsidiary assets. Same question, different platform, depending only on which engine the buyer opened.

Integration was the second. ChatGPT named Tenable and Bitsight as the vendors for connectivity with ticketing, workflow and analytics tools. Google AI Overview named IONIX and CyCognito for the same criterion.

The third disagreement is the edge of the category itself. Microsoft Copilot alone returned 15 product entries no other engine named, among them Qualys VMDR and ServiceNow Vulnerability Response. Perplexity alone returned 7, among them Bugcrowd Savant Vista and CATAAM. Google AI Mode and Gemini each held 14 entries of their own. Ask about attack surface management and some engines answer with vulnerability management and security operations tools.

Product naming still splits visibility, even after we merged plan and edition spellings of one product. The corpus held 124 product entries: Rapid7 appeared as 10 separate entries, Palo Alto Networks as 9 and Tenable as 8. Of the 124, 69 recorded exactly 1,000 mentions and 81 were named by a single engine. Some of that is real product breadth. Some is naming drift across a vendor's own published material, and only consistent naming lets a measurement consolidate it.

What Should Vendors and Buyers Do With This?

Buyers should treat the AI shortlist as one input. It is narrower than an analyst report, ordered differently by engine, and the vendor named first differs by engine with that engine's evidence base, which is not the same as the best fit for your environment. Where the engines actively disagree, as they do on post-acquisition discovery, evaluate both named platforms.

Vendors have five things to act on.

Track three numbers, not one. Mention volume, first-mention share and mean position answered differently in this category, and they do not respond to the same work.

Treat your own site as the primary citation surface. 51.3% of everything the engines cite sits on a vendor domain. Documentation, product and integration pages are what most engines read, with analyst and review pages second at 11.1%.

Instrument at least three engines from different families. With mean overlap at 0.16 and Microsoft Copilot below 0.12 with everyone, a program tuned to ChatGPT tells you almost nothing about Copilot. The manual version is to run every prompt across every engine and log what each one cites; at scale, one of the 15 AI search monitoring tools we have compared does it for you.

Name products consistently. Rapid7's ten entries and Palo Alto Networks' nine are partly product breadth and partly drift. Only the vendor can tell which.

Skip market-by-market programs for these four markets. The leading group did not change, and only Germany changed its order. Two things to stop reporting as levers: link decay, at 0.4%, and sentiment, which sat between 0.69 and 0.84 for every leading brand with a recorded score, a spread too narrow to separate any two of them.

Frequently Asked Questions

How many AI responses did this attack surface management study analyze?

240,000 responses, generated from 10,000 enterprise buyer-intent queries issued to six AI engines across four markets in August 2026. The responses contained 1,172,000 cited sources and named 124 distinct product entries.

Which ASM vendor has the highest AI search visibility?

CyCognito, with 160,000 mentions. Palo Alto Networks followed at 157,000 and Microsoft at 115,000. Seven brands were named by all six engines tested: those three plus Tenable, CrowdStrike, IONIX and Axonius.

Which vendor do AI engines name first for attack surface management?

It depends on the engine. CyCognito opened the response in Gemini, Google AI Mode and Google AI Overview. Palo Alto Networks opened ChatGPT and Microsoft Copilot, and was read earlier on average in four of the six engines. Microsoft opened Perplexity.

Do all AI engines cite the same sources for security vendor recommendations?

No. Cited-domain overlap averaged 0.16 across the 15 engine pairs. The highest value, 0.38, was between Google AI Mode and Google AI Overview, two surfaces from the same company, and no pair from different vendor families exceeded 0.20.

What percentage of AI-cited sources are dead links?

0.4% across the full corpus of 1,172,000 cited URLs. The rate ranged from 0.0% for Microsoft Copilot, Gemini and Perplexity to 0.6% for ChatGPT and Google AI Mode.

Final Thoughts

Six AI engines answering the same attack surface management questions converged on a narrow vendor shortlist, disagreed on which vendor to name first, and drew that agreement from evidence bases that barely intersect but share one class of source: the vendors' own sites. For vendors, the actionable gap is measurable. It is the distance between being in the shortlist and opening it, and it has to be closed once per engine.

We have seen the same shape elsewhere in this series, in AI engines picking the same three DSPM vendors and in Vanta leading compliance automation while the engines cite almost none of the same sources. Full per-engine tables, the source-type breakdown, the overlap matrix, and the study's validity limits are available in the complete benchmark report.

Revision note: the corpus behind this post was reprocessed under corrected citation-counting and product-naming rules. Citation totals, dead-link rates, source overlap and product-entry counts changed materially from the August edition, and the figures above supersede it. The response corpus itself is unchanged.

Disclosure: GrackerAI publishes this research and sells AI search visibility measurement for the category it covers. The study measured how AI engines describe vendors. It did not test, evaluate, or rank any vendor's product, and no vendor paid for or requested placement.