Not one of them does. We fetched the homepages of 26 agentic-product vendors with 9 crawler identities: all 26 serve the buyer-facing answer-surface crawlers (ChatGPT, Claude, and Perplexity fetchers among them), 23 of 26 actively market to AI agents on their own domains, and the number that do both while denying the agent channel is zero. The prevalent "software vendors are walling themselves off from AI" story is, at the network edge for this cohort, the opposite of what the wires say.
The one split that matters: assist, don't train
Only two vendors block the usual training crawlers at the edge, and both serve 200 to every answer-surface crawler: Zoho CRM (blocks GPTBot, ClaudeBot, CCBot) and Botpress (same trio). That is a coherent policy rather than a contradiction: sell to agents, decline to be their training data. On the robots.txt side, Attio names GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Google-Extended, and PerplexityBot with explicit allow groups while block-all-ing exactly one crawler, CCBot. "Welcome the answer engine, refuse the scraper" is a deliberate, legible strategy at the leaders.
| Vendor | Markets to AI agents | Serves answer-surface crawlers | Blocks at the edge | Notes |
|---|---|---|---|---|
| ActiveCampaign | yes | all 5 | none | — |
| Attio | yes | all 5 | none | robots block-all: CCBot |
| Bland | yes | all 5 | none | — |
| Botpress | yes | all 5 | CCBot, ClaudeBot, GPTBot | — |
| Braze | yes | all 5 | none | — |
| Close | yes | all 5 | none | — |
| Customer.io | yes | all 5 | none | — |
| Ellipsis | yes | all 5 | none | — |
| GMass | yes | all 5 | none | — |
| HubSpot | yes | all 5 | none | — |
| Klaviyo | yes | all 5 | Google-Extended | anti-spoofing bot-fighting suspected (Google-Extended) |
| Lindy | yes | all 5 | none | — |
| Mailchimp | yes | all 5 | none | — |
| Make | yes | all 5 | challenge wall (all bots) | — |
| Mem | yes | all 5 | none | — |
| Mixmax | yes | all 5 | none | — |
| Omnisend | yes | all 5 | Google-Extended | anti-spoofing bot-fighting suspected (Google-Extended) |
| Pipedrive | yes | all 5 | Google-Extended | anti-spoofing bot-fighting suspected (Google-Extended) |
| Relevance AI | yes | all 5 | none | — |
| Zapier | yes | all 5 | none | — |
| Zoho CRM | yes | all 5 | CCBot, ClaudeBot, GPTBot | — |
| folk | yes | all 5 | none | — |
| n8n | yes | all 5 | none | — |
| Mailmeteor | no | all 5 | none | no agent-facing marketing found |
| Motion (motion.app 301s to usemotion.com) | no | all 5 | none | no agent-facing marketing found |
| Yesware | no | all 5 | none | no agent-facing marketing found |
Reading the labels
"Markets to AI agents" is our reading of the vendor's own marketing copy: an agent store, an agentic feature, or copy addressed to autonomous buyers. Quotes are captured verbatim from vendor-owned pages. "Serves answer-surface crawlers" refers to the five buyer-facing fetchers: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, and PerplexityBot.
Two labels exist because the naive reading lies. Three vendors (Klaviyo, Omnisend, Pipedrive) return 403 to Google-Extended, but they also 403 a plain fake-Googlebot user-agent while a normal browser agent passes: that is anti-spoofing bot-fighting, not AI policy, so we label it separately. Make serves a Cloudflare challenge to every identity including our control fetch: a vantage artefact, not a crawler policy. Our own tooling initially mislabelled two vendors here; the parser sealed robots groups only on Disallow and swallowed allow-only welcome groups. The bug is fixed, regression-tested, and every affected benchmark was re-run, because a scan that cannot catch its own parser lying is worse than no scan.
Method
- Fetch each vendor homepage with each crawler's exact user-agent string; record HTTP status.
- Parse each robots.txt for per-crawler groups; record allow/disallow verdicts.
- Fetch each homepage once with our own research user-agent as a control; a status equal to the control means the wall is not crawler-specific.
- Fetch plain 'Googlebot' as a second control; vendors that block it are running anti-spoofing bot-fighting, which we label separately from AI policy.
- Re-scan weekly; a verdict counts only after it holds across consecutive scans (2 so far).
Vantage: TestedActually build machine, datacenter IP, 2026-10-06. We publish the vantage so results can be reproduced independently. Between the first two weekly scans, 232 of 234 cells were identical; the two differences were transient network errors that read 200 on re-scan, with robots verdicts unchanged throughout.
Limitations
- Edge and robots.txt only: JavaScript rendering, paywalls, and login walls are separate surfaces we have not measured here.
- Homepage only: a vendor could allow crawlers to the homepage and deny deeper paths.
- One vantage point; WAF behaviour varies with IP reputation.
- Two weekly scans so far. A verdict only counts once it holds across consecutive scans, and the scan continues weekly.
- "Markets to AI agents" is our reading of vendor copy, captured verbatim so you can check it.
Reproduce
The scanner, the joined classification, and the verbatim marketing quotes live in the site's repository, one JSON per weekly scan, no manual edits. The machine-readable version of this page is at /research/agent-denial.json.