How AI assistants answer B2B buyer questions in AI engineering and security SaaS

Summary

We covered two categories: AI engineering (agent observability and evaluation, AI gateways, agent infrastructure) and security (compliance automation and GRC, AI security, AI-assisted security operations, application security, secrets management). We also ran a smaller set of questions through the model APIs to compare. Five things stood out:

1. Funded challengers are often absent from their own category’s answers. 7 of the 25 companies we screened were named in none of their consumer-app answers, and 3 were named in all of them.
2. “How do I…” questions often get no vendor at all. Problem-phrased questions produced no vendor in 5 of 10 app answers. The same topics phrased as “Which tools…” produced a vendor every time (0 of 16 app answers without one).
3. On an “alternatives” question, the cited evidence was comparison pages. All 8 inline sources ChatGPT cited across two runs were comparison or “alternatives” pages, not vendor product pages.
4. The apps picked up a new page within two days; the APIs didn’t. A vendor’s new comparison page was used by Perplexity and Google AI Mode two days after it went live. No API answer that day named the vendor or cited its site.
5. Cheap API models are not a stand-in for the apps. Logged-out ChatGPT named the challenger on 2 of 6 questions. The cheaper API ChatGPT model named none of the six challengers in 13 answers.

How to read the numbers

• “Named in X of N” counts answers whose text recommends or lists the company. A citation or source chip alone doesn’t count as named.
• Every number says whether it comes from the consumer apps (what a buyer sees in the browser) or the API (a model called directly by code), with the date and location. App results lead because that is what buyers see.
• 95% Wilson intervals are given where N is 10 or more. Treat them as rough ranges: answers to the same question are not fully independent, so the true uncertainty is wider.
• Companies are anonymised as Company A, B, C… Only categories are named. Well-known category leaders are named where they help explain an answer.

Finding 1: Funded challengers are often absent from their own category’s answers

7 of 25 companies were named in 0 of their app answers (95% Wilson 14–48%). 3 of 25 were named in all of them.
Apps, 18 Oct 2026, India.

• Each company was checked on 1–2 buyer questions about its own category, 3–7 app answers per company, across logged-out ChatGPT, Perplexity and Google AI Mode.
• Example (AI engineering). Company A was named in 0 of 6 app answers on two questions describing its core use case, including a question that describes its headline feature. On one of those questions, LangSmith, Langfuse and Arize were named in all 3 app answers.
• Example (security). On an “AI-agent GRC platform” question, Company L was named in 1 of 3 app answers. The only mention came from a signed-in Google session (weak evidence, see limits). Logged-out ChatGPT listed enterprise GRC incumbents in both runs.
• Where challengers did well. In AI engineering infrastructure, logged-out ChatGPT often named funded newer companies first or second: for agent memory, embedded vector databases, document parsing and web data for agents. Several were dropped from the screen after a single check for that reason, so they are not all in the table. Absence is common, but it isn’t the rule.
• By assistant (descriptive only; each assistant saw a different mix of companies): ChatGPT logged out named the company in 16 of 53 answers, Perplexity in 15 of 28 and Google AI Mode in 8 of 21.

Why this matters: shortlists from AI assistants are short. A company missing from its own category answer is missing from the list a buyer starts with.

Finding 2: “How do I…” questions often get no vendor at all

Problem-phrased questions: no vendor named in 5 of 10 app answers (95% Wilson 24–76%).
Tool-asking questions: 0 of 16 app answers without a vendor (95% Wilson 0–19%).
Apps, 18 Oct 2026, India.

• The 10 app answers came from four problem-phrased questions, such as giving an AI agent access to email, chat and code-hosting tools, or running AI-generated code safely.
  • ChatGPT (logged out) named no vendor in 4 of 6. It answered with build-it-yourself architectures instead: an MCP client and tool gateway, or containers and microVMs.
  • Perplexity (one run, signed-in account) did the same once.
  • Google AI Mode named vendors in all 3 of its answers.
• The API shows the same effect at larger N:
  • “How do I detect silent failures in AI agents in production?” named no vendor in 10 of 11 API answers (95% Wilson 62–98%).
  • “How can AI automate third-party vendor risk assessments?” named no vendor in 9 of 11 (52–95%).
  • Rewritten as “Which tools detect silent failures in production AI agents…?”, the first question named no vendor in 0 of 8.
  • Across six questions rewritten to ask for tools, only 1 of 48 API answers named no vendor (95% Wilson 0.4–11%).

Why this matters: the earliest-stage buyer question, asked before the buyer knows the category name, is often answered with a DIY recipe. A vendor that wants that buyer needs to be the example inside the recipe: documentation, code samples and how-to guides that an answer can draw on.

Finding 3: On an “alternatives” question, the cited evidence was comparison pages

8 of 8 inline sources were comparison or “alternatives” pages.
ChatGPT app, logged out, 18 Oct 2026, India; 1 question, 2 runs; N too small for an interval.

• The question asked for alternatives to the best-known SOC 2 compliance platform for early-stage startups. Across two runs ChatGPT cited 8 distinct pages; one appeared in both runs.
• Vendor websites appeared only as “Website:” links under each recommendation, not as the source for any claim.
• Only 1 of the 8 cited pages was clearly a vendor’s own page. We couldn’t quickly attribute the others, so we make no claim about who owns them.
• The API told a different story. In an earlier API sample, the ChatGPT-family API model cited vendors’ own product, pricing and demo pages. The app and the API rely on different evidence, which is Finding 5 again.

Why this matters: in this answer, the pages deciding the shortlist were comparison pages, not product pages. Being included, accurately, in the comparison pages an assistant already cites can matter as much as a vendor’s own site.

Finding 4: The apps picked up a new page within two days; the APIs didn’t

Vendor named in 2 of 4 app answers (Perplexity logged out 1 of 1; Google AI Mode signed in 1 of 1; ChatGPT logged out 0 of 2) vs 0 of 8 API answers to the same question that day.
The vendor’s domain was cited in 0 of 48 API answers that day (95% Wilson 0–7.4%). 18 Oct 2026, India.

• The question asked which tools detect silent failures in production AI agents.
• An AI engineering vendor had published a “best tools, compared” page two days earlier.
• In the apps, Perplexity named the vendor and cited its site. Google AI Mode named it first and cited the new page itself.
• No API answer to the same question named the vendor or cited its domain.
• The vendor is not identified here. It is one of the companies in Finding 1, and naming the page would point to it.

Caveats that matter here:

• The API answers were collected about 10 hours before the app answers, so timing and channel are confounded.
• The Google session was signed in and personalised.
• The page is the vendor’s own, so it ranks itself well. The notable part is the speed of pickup, not the ranking.
• This is a single observation. We re-run the same questions on 15 November and publish the result, whatever it shows.

Why this matters: new content can reach some assistants in days, not model generations. Measuring only through APIs would have missed it.

Finding 5: Cheap API models are not a stand-in for the apps

ChatGPT app (logged out) named the challenger on 2 of 6 questions and ranked one first (apps, 18 Oct 2026, India; 1 run each).
The cheaper API ChatGPT model named none of the six challengers in 13 answers (API, 18 Oct 2026, no location; 95% Wilson 0–23%).

• Brand overlap tells the same story. Comparing the brands each source named, the app and the API model shared 0.37 on average (Jaccard index). Two API runs of the same question share 0.63–0.66.
• The same pattern holds for Perplexity. On the silent-failures question, the Perplexity app named three challengers. The Perplexity API model (Sonar) named none of them in 3 runs each.
• All disagreements ran the same way: the API missed companies the apps named. An API “named” result was usually confirmed in the apps. An API “absent” result needed an app check.

Why this matters: a visibility audit built on cheap API calls can report false gaps and miss recent wins. We use APIs for screening and source analysis only. Client numbers come from repeated checks in the consumer apps, as set out in our methodology.

We measured ourselves too

Using the same method on five questions about our own field (“AEO agency for B2B SaaS”, “how to measure AI visibility”, “AI crawler checker tool”, and two others), aeoledger.com was named or cited in 0 of 15 answers (ChatGPT and Perplexity logged out, Google signed in with personalisation off; 18 Oct 2026, India; 95% Wilson 0–20%). Google also corrected a search for our brand name to a common phrase. The site went live that week. We re-run the same questions on 15 November and publish the result, whatever it shows.

Method

Fieldwork

18 Oct 2026. API samples in the early hours (IST); app checks from about 13:20 to 20:00 IST. Bengaluru, India.

Questions

Written in buyer wording, without the target company’s name. Some “alternatives” questions name a category leader; that leader is excluded from counts on that question.

Consumer apps

ChatGPT logged out (default model, web search as chosen by the app), a fresh chat per question. Perplexity logged out for the first checks; its logged-out query limit was then reached, and later checks used a signed-in free account with a fresh thread per question. Google AI Mode in a signed-in browser showing “Personalising”. The own-site check used Google Search’s AI Overview with personalisation switched off. Mostly one run per question per assistant; a few questions had 2–3 runs.

APIs

Cheaper tiers of the ChatGPT, Claude and Perplexity model families, with web search on; no system prompt, no location. 2–3 runs per question per model. Gemini was in the earlier API sample only.

Scoring

Answers read by hand (apps) or matched against a brand list (API). “Named” means recommended or listed in the answer text. Citations are recorded separately. A brand named in the question itself is excluded from that question’s counts.

Companies

25 venture-funded companies in the two categories with at least 3 app answers each. Smaller checks (1–2 answers) are excluded from Finding 1.

Intervals

Wilson score interval at 95%,^[1] recommended over the simple ± formula for small samples.^[2] Shown only for N of 10 or more.

Limits

1. Not a random sample. The companies were screened as possible outreach prospects, and some as category benchmarks. The 7-of-25 figure describes this screen, not all challengers in these categories. Pooled answer counts mix companies, so they come without intervals.
2. Few runs. Most app cells are a single run. Two runs of the same question share only about two-thirds of the brands they name, so any single answer can miss or include a company by chance.
3. Signed-in sessions. Google AI Mode was signed in and personalised throughout. Part of the Perplexity data came from a signed-in account. Both could tilt answers toward pages that account had visited.
4. One location, one day. All app checks ran from India on 18 Oct 2026. Assistants localise and change often; buyers in the US or Europe may see different answers.
5. Claude was checked through the API only (no clean app account), and Google AI Overviews only in the own-site check.
6. Hand reading. App answers were read and tallied by hand, without screenshots. Brand lists may be incomplete.
7. Timing confound in Finding 4. API and app answers were about 10 hours apart.
8. Interval assumptions. Wilson intervals assume independent answers; repeated answers to one question are correlated, so the real uncertainty is larger.

Data and sources

Download the data (CSV, 45 rows): one row per number on this page, with finding, series, named count, N, percentage, 95% Wilson interval where shown, source (apps or API), assistant, date, location and notes. Companies are anonymised. Licensed CC BY 4.0: reuse it with a link to this page.

1. Wilson, E. B. (1927). “Probable inference, the law of succession, and statistical inference.” Journal of the American Statistical Association 22(158): 209–212.
2. Brown, L. D., Cai, T. T., DasGupta, A. (2001). “Interval estimation for a binomial proportion.” Statistical Science 16(2): 101–133.

Next update: we re-run the same questions on 15 November and publish the result, whatever it shows.

How to cite and changes

AEOLedger. “How AI assistants answer B2B buyer questions in AI engineering and security SaaS.” Research v1.0, 19 October 2026. https://aeoledger.com/research/ai-answers-oct-2026/

• v1.0, 19 October 2026: first published.

Spotted an error? Write to hello@aeoledger.com. Corrections are dated here.

Read our research often? Add AEOLedger as a preferred source on Google, so our pages show up more in your Google results.