AgentGate

1. One-liner

Fires the moment your firewall starts turning away ChatGPT’s shopper — the channel that converts 42% better.

2. Trend signal — why now?

Three things collided in the last ninety days, and the third one has a date on it.

AI-referred traffic stopped being a novelty and became the best traffic you get. Adobe’s Q1 2026 data: AI-sourced traffic to US retail sites grew 393% year over year, and in March 2026 it converted 42% better than non-AI traffic. Twelve months earlier that same channel converted 38% worse — an 80-point swing in a year. Revenue per visit from AI referrals now runs 37% above everything else. On Shopify specifically, Q1 2026 AI-driven traffic grew 8× and orders from AI search grew nearly 13×.

Meanwhile, roughly a quarter of sites are quietly blocking the bots that carry it. The blocking almost never happens in robots.txt — it happens at the edge, where nobody looks. Cloudflare’s Bot Fight Mode is on by default on the free tier and classifies AI crawlers as unwanted bots. Wordfence has no AI-bot allowlist and its rate-limiting heuristics throttle GPTBot and ClaudeBot the same way they throttle scrapers. WAF rate rules tuned tight serve a JavaScript challenge on request four; AI crawlers don’t run JavaScript, so they see a brick wall and go cite a competitor instead. Cloudflare’s own community forum has site owners reporting AI crawlers still returning 403s that don’t even appear in security events — with every bot-protection toggle already set to allow.

And on 15 September 2026, Cloudflare flips the default against you. New customers, new sites from existing customers, and all existing free customers get Training- and Agent-class bots blocked by default on any page that carries an ad. “Agent” explicitly includes ChatGPT’s fetch bot and browser-driving agents. Cloudflare fronts 23.4% of all websites. There is an opt-out, and the overwhelming majority of small merchants will not know it exists until their AI channel is already gone.

That’s the shape of it: your highest-converting traffic source, a silent failure mode at a layer your marketing team can’t see, and a hard date when the default gets worse.

Provenance:

• Signal 1 (demand): AI traffic to US retail sites +393% YoY in Q1 2026, converting 42% better than non-AI traffic in March 2026; revenue per visit 37% higher — https://techcrunch.com/2026/04/16/ai-traffic-to-us-retailers-rose-393-in-q1-and-its-boosting-their-revenue-too/ — 2026-04-16
• Signal 2 (feasibility/pain): ~27% of sites unintentionally block AI crawlers via CDN/WAF/plugin rules that robots.txt never sees; Cloudflare community reports of invisible 403s to ClaudeBot/GPTBot despite bot protection disabled — https://suganthan.com/notes/firewall-blocks-ai-crawlers/ and https://community.cloudflare.com/t/ai-crawlers-still-blocked-unsuccessful-despite-bot-fight-mode-off-block-ai-bots/927107 — 2026-05
• Signal 3 (economic): Cloudflare blocks Agent + Training bots by default on ad-carrying pages from 15 Sept 2026 for new customers and all existing free customers; AI-visibility tooling raised $300M+ between summer 2025 and spring 2026, Profound alone at a $1B valuation with 700+ enterprise customers — https://techcrunch.com/2026/07/01/cloudflares-new-policy-pushes-ai-companies-to-pay-for-publishers-content/ — 2026-07-01 Category: Platform shift

3. The opportunity

The AI-visibility category raised $300M+ and every dollar of it went to answering “are we mentioned?” Profound ($499/mo, 700+ enterprise customers, $1B valuation), Peec (€89–199/mo), Otterly ($25/mo) all track prompts, citations, and share-of-voice across ChatGPT and Perplexity. That’s the relevance problem.

Nobody is selling the access problem, and access is upstream of relevance. If GPTBot gets a 403 at your edge, no amount of citation engineering matters — you are not in the index at all, and the ranking dashboard just shows a flat line you’ll misread as “we need better content.”

The closest things that exist:

• Free one-shot checkers (aicrawlercheck.com, Suganthan’s AI Crawler Access Checker). Type a domain, get a score, done. No monitoring, no alerting, no re-check when someone updates a plugin next Tuesday. A snapshot of a thing that changes silently is close to worthless.
• Peec’s Crawl Insights / Profound’s Agent Analytics — real, but passive log analysis. You upload server logs, it tells you what already happened. It’s an autopsy, not a smoke alarm. It also requires you to have log access and a €89–499/mo brand-visibility subscription you bought for a different reason.

The gap is an active prober: something that continuously hits your site as GPTBot, ClaudeBot, PerplexityBot and a browser-driving agent, from real crawler-like network paths, and screams the same day the answer changes from 200 to 403. Plus, for merchants, the second-order question nobody checks: even when the agent gets in, can it actually complete the job — find the price, read the stock status, reach a checkout that doesn’t demand a JS-only flow?

Incumbents won’t rush here. Profound sells six-figure enterprise deals to CMOs who want share-of-voice charts; a $39/mo uptime-style alarm for a Shopify merchant is a rounding error that would drag their ACV down. This is a classic too-small-for-them wedge with a hard deadline attached.

4. Target market

• Primary customer: Owner or head-of-ecommerce at a DTC brand doing $500K–$20M/yr on Shopify/WooCommerce/BigCommerce, running Cloudflare (usually free tier) or a WordPress security plugin, with 1–3 people on marketing and no dedicated infra person. Secondary: the GEO/AEO agencies now charging $2,000–$8,000/mo retainers, who need a monitoring layer to put under those retainers.
• Why they buy: They just watched AI become their best-converting channel and they have no instrument that tells them if it’s still connected. The pain is specifically silent — as one write-up puts it, “you only discover the problem when your competitors start getting cited in ChatGPT for queries you should own.” Nobody wants to find out in November that the channel died in September.
• Rough TAM reasoning: Cloudflare fronts 23.4% of all websites; Shopify alone has millions of merchants and turned on Agentic Storefronts by default for US sellers in March 2026. The serviceable slice — merchants with enough revenue to care and enough AI traffic to lose — is realistically in the low hundreds of thousands globally. I need 500 of them at $79/mo to clear $470K ARR. The market is not the constraint here.
• Why now for them: 15 September 2026. A dated, externally-imposed change to a default they didn’t choose, affecting a channel they just learned to value. That is the single most sellable thing in B2B software.

5. Product sketch (MVP)

• Live access probe. Every few hours, fetch your key URLs while presenting as GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended and a browser-driving agent. Record the real status code, the challenge type, and whether usable HTML came back or a JS stub.
• Same-day block alert. Email/Slack the moment a 200 turns into a 403, a 202 challenge, or a rate-limit wall. This is the whole product in one sentence: your AI channel just went dark, here’s the exact rule that did it.
• “Who did it” attribution. Name the layer — Cloudflare Bot Fight Mode, a WAF rate rule, a Wordfence heuristic, a robots.txt line an SEO plugin rewrote — and give the specific setting to change. Diagnosis without the fix is a support ticket, not a product.
• September 15 readiness check. One-click audit: are you on Cloudflare, are you free-tier, do your money pages carry ad tags, are Agent-class bots about to be blocked by default, and have you set the opt-out? Runs as a standing check, not a one-time scan.
• Agent task simulation. Can an agent actually transact? Drive a headless agent through find-product → read price → read stock → reach checkout, and report where it stalls (JS-only price, missing JSON-LD, no product feed, cart behind a challenge).
• Change diffing. Watch robots.txt, meta robots, X-Robots-Tag and CDN behaviour for changes, and tell you which deploy or plugin update broke it.
• Agency multi-site view. One dashboard, N client domains, one weekly “all clear / 2 sites blocked” digest they can forward to clients.

6. AI angle — what’s load-bearing

Honest answer: the AI here is the subject, not the engine. The core loop is disciplined network probing and status-code diffing — that’s crawling and infrastructure work, not model work. If I claimed an LLM was doing the heavy lifting I’d be decorating a form, and this repo has a rule about that.

Where models genuinely earn their place:

1. Agent task simulation — driving a real browser-use agent through a purchase path and deciding “did this succeed?” is a judgement call over messy, rendered pages. That is a vision/agent task and there’s no deterministic way to fake it.
2. Fix attribution — mapping an observed edge behaviour (a 202 with a JS stub, a 403 absent from security events) to the specific vendor setting that caused it, across Cloudflare, Wordfence, Sucuri, AWS WAF and a dozen Shopify apps, is fuzzy pattern-matching over a long tail of configurations. An LLM over a growing corpus of observed fingerprints beats a hand-written rules table, and it compounds.

Strip the AI out and you still have a useful monitor — a worse one, with no transaction testing and generic advice. So I’m scoring this honestly rather than pretending the AI is load-bearing when it’s load-sharing. What makes the product timely is that the customer’s revenue now depends on AI agents; that’s the real dependency.

7. Localization angle (if any)

N/A — this is a global play. The blocking layer (Cloudflare, WAF plugins) and the agents (OpenAI, Anthropic, Google, Perplexity) are the same everywhere. Pricing in USD, sold in English to merchants who already buy USD SaaS. Forcing an India-first angle here would shrink the market for no gain — the Adobe/Shopify conversion data that makes this urgent is US retail data, and that’s where the willingness to pay is concentrated. Sell where the money already moved.

8. Business model — path to $1M–$5M ARR

• Pricing: $39/mo single site (up to 25 monitored URLs). $79/mo brand tier (transaction simulation + change diffing). $249/mo agency tier (up to 25 client domains, white-labelled weekly digest).
• ACV: ~$950 blended, assuming the mix skews to the $79 brand tier with a meaningful agency tail.
• Rough math to $1M ARR: 700 brand accounts at $79/mo = $663K, plus 120 agency accounts at $249/mo = $359K → $1.02M. That’s ~820 paying accounts. Plausible for a two-person team inside 18 months given the deadline-driven demand.
• Rough math to $5M ARR: needs the agency channel to carry it — ~1,000 agencies at $249/mo is $3M, and agencies resell into their own $2K–$8K/mo retainers, so the tool becomes a line item they mark up rather than a cost they resent. Add an enterprise tier ($999/mo, unlimited domains, API, SSO) for the mid-market brands who’ll want it once this is a standing board metric.
• Expansion path: URLs monitored → sites monitored → transaction paths simulated. Natural upsell into “agent readiness” remediation reports. The honest ceiling is that this is a monitor, and monitors are cheap; the way ACV grows is by widening what’s monitored, not by charging more for the same alarm.
• Margins: trivially good. This is HTTP requests and a modest amount of headless-browser time. Cost per account is a couple of dollars a month at worst.

9. Go-to-market wedge — first 100 customers

The deadline does the selling. I don’t have to explain why they should care; I have to reach them before 15 September.

1. The free scanner is the funnel, and it’s an ad. Ship Is your site blocked? — enter a domain, get a real probe from six agent user-agents in ten seconds, see the actual status codes. Free checkers already exist and get traffic, which proves the demand; the difference is mine ends with “this is a snapshot. It was fine last month too. Get told the day it changes — $39/mo.” Conversion of free-scan → paid on a fear-driven monitoring product realistically runs 1–3%; I need ~5,000 scans to get the first 100 paying.
2. Scan first, then cold-email with the finding. Take the Shopify/BigCommerce app-store and BuiltWith lists, filter to Cloudflare-fronted stores, and run the probe before I email. The email isn’t a pitch, it’s a result: “GPTBot got a 403 from yourstore.com on these 4 URLs on 14 July. Here’s the Cloudflare setting that did it.” A cold email containing a specific, verifiable, embarrassing fact about the recipient’s own site converts unlike anything else — I’d expect 8–12% reply on 2,000 sends, and a good chunk of those are already-broken sites, which are the easiest possible close. This alone should produce the first 100.
3. Sell the deadline to the agencies, not the merchants. ~50–100 GEO/AEO agencies charging $2K–$8K/mo retainers need to show clients they’re on top of the September change. Offer the agency tier free until 15 September in exchange for putting client domains in. Each agency brings 10–25 domains. Twenty agencies = a few hundred monitored sites and a recurring channel that doesn’t depend on me.
4. Own the “September 15” search term. Publish the definitive plain-English page on what Cloudflare’s default flip does to ecommerce, with the scanner embedded in it. This is a dated event that every merchant and agency will search for in August. Two months of head start on one keyword with a hard deadline behind it is a real, if temporary, distribution asset.

10. Build complexity — justification

Low. The MVP is a scheduler, a pool of outbound requests with controlled user-agents and honest identification, status-code diffing, and an alerting layer — a competent solo builder ships that in 5–6 weeks. The two genuinely fiddly parts: (a) probing from network paths that behave like real crawler traffic rather than a datacenter IP that every WAF already knows, and (b) the headless-agent transaction simulation, which is the one piece that will need real iteration. Neither is research-grade. Off-the-shelf headless browsers, off-the-shelf agent frameworks, standard web stack.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Probing sites you’re authorised to monitor, at the owner’s request. Crawlers identify honestly; this is not evasion. Domain ownership must be verified at signup — that’s a hard requirement, not a nicety.
Ethical — no harm / dark patterns	✅	Fear-adjacent marketing, so the finding must always be true and verifiable. Show the raw status code and let it speak. No invented urgency — the September date is real and doesn’t need embellishing.
Market exists (evidence above)	✅	Adobe conversion data, ~27% accidental-block rate, $300M+ into adjacent tooling, Cloudflare’s dated default change.
1–5 person team can build this	✅	One or two people.
Launchable with <$50K / ₹40L	✅	Realistically <$10K. Proxies, headless browser time, hosting.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Real money, and the failure is silent — the worst kind. Docked because it’s invisible pain: a merchant losing AI traffic often doesn’t know it, so the urgency has to be manufactured by me showing them the 403. Hair-on-fire only after the scan.
Demand evidence	15	13/15	Strong and independent: Adobe’s conversion numbers, the ~27% block rate, forum threads of owners hitting invisible 403s, $300M+ funding into adjacent AI-visibility tools, and a dated platform change. A skeptic nods.
Build feasibility	15	14/15	5–6 weeks solo. Nothing here is hard except making the probes look like real crawler traffic.
Distribution clarity	15	12/15	The scan-then-email motion is genuinely excellent — a verifiable fact about the recipient’s own site. Docked because free-scanner→paid conversion on a $39 product is a numbers game needing real volume, and I’m guessing at the rate.
Revenue mechanics	15	11/15	Pricing is benchmarked (Otterly $25, Peec €89) and margins are excellent, but $39–79/mo means ~820 accounts for $1M. That’s a lot of small logos to acquire and keep. The agency tier is what makes the math work, and it’s the least proven part.
Time to first revenue	10	9/10	Pre-sellable before the build finishes. Scan 200 sites, email the broken ones, take money for the alert. Weeks, not months.
Defensibility	10	2/10	This is the problem. A weekend clone. No network effects, no proprietary data on day one, no regulatory moat. Peec or Otterly could ship active probing as a feature next quarter.
Total	100	77/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · content-heavy

Technical because the probing infrastructure and agent simulation are the product. Content-heavy because the free scanner plus the “September 15” keyword is the distribution strategy — someone has to write the definitive page and make it rank before August.

Key assumptions to validate (3–5)

1. Assumption: A meaningful share of real merchant sites are actually blocking AI agents right now — the ~27% figure holds on a list I pick myself, not just in a vendor’s blog post. How to test: Build the prober first, before anything else. Run it against 500 Shopify/BigCommerce stores from a BuiltWith export. If fewer than 10% show a block or challenge to GPTBot, the cold-email wedge collapses and so does the product.
2. Assumption: Merchants will pay a recurring fee for an alarm rather than pay once for a fix. How to test: Email 200 merchants whose sites I’ve confirmed are blocked. Offer two things at the same price: a one-time fix report, and a $39/mo monitor. If they overwhelmingly take the one-time report, this is a $200 consulting product, not SaaS — and I should know that in week two.
3. Assumption: Agencies will adopt this as a monitoring layer under their $2K–$8K retainers. How to test: Get 10 GEO/AEO agencies on calls before writing the agency tier. Ask directly whether they’d white-label it and what they’d pay. If they say “we’d just build this,” believe them.
4. Assumption: The September 15 change actually bites ecommerce sites, not only ad-carrying publishers. How to test: Read the final Cloudflare policy carefully and probe real merchant sites after the flip. The default targets pages “that host ads” — a DTC store with no ad tags may be untouched, which would blunt the deadline. This is the single most load-bearing assumption in the entire proposal and it deserves the most scrutiny.

Risk flags

1. Defensibility (severe): 2/10 and I’m not dressing it up. Active probing is a feature, not a company. The only durable asset is the corpus of observed edge-fingerprints → causes, which compounds if volume arrives fast. If Peec or Profound bolts on active probing before that corpus matters, the differentiation evaporates and this becomes a price fight I lose.
2. Platform dependency (severe): The entire “why now” is one company’s configuration default. If Cloudflare softens the 15 September change, exempts ecommerce, or ships a merchant-friendly one-click “allow shopping agents” toggle in its own dashboard — which is exactly the kind of thing Cloudflare ships — the urgency evaporates overnight. Cloudflare already has AI Crawl Control. They are one product decision away from making most of this free.
3. Market timing (moderate): This is a window, not a market. The window opens now and closes when either (a) the platforms make agent access boringly reliable, or (b) accidental blocking becomes a solved default. Two years, maybe three. Build accordingly — take the money, don’t plan a decade.
4. Ethical/positioning (moderate): Fear-driven monitoring products slide easily into FUD. The discipline is that every alert must correspond to a real, reproducible status code, and the free scanner must be honest when a site is fine. The moment the scanner starts crying wolf to drive conversions, the brand is dead and it deserves to be.

14. Structured verdict

Score:                  77/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical solo founder who can write. Needs to ship a prober fast and own a keyword before August.
Time to revenue:        3–6 weeks (pre-sellable off the scan results before the product is finished)
Capital to launch:      <$10K / ₹8L
Top 3 assumptions to validate first:
  1. Real block rate on 500 real merchant sites is ≥10% — build the prober and run it before building anything else
  2. Merchants buy the recurring alarm, not a one-time fix — A/B the offer against 200 confirmed-blocked sites
  3. The Sept 15 default actually bites ad-free ecommerce sites, not just ad-carrying publishers — read the final policy, probe after the flip
Kill criteria:
  - Abandon if <10% of 500 probed merchant sites are blocking or challenging AI agents
  - Abandon if Cloudflare ships a one-click "allow shopping agents" default before launch — that's the whole product, free, from the layer that owns the problem
  - Abandon if confirmed-blocked merchants take the one-time fix over the subscription by >4:1
  - Abandon if free-scan → paid conversion is <0.5% after 5,000 scans

15. Next step — 1-week validation sprint

• Day 1–2: Build only the prober. Six user-agents, real status codes, honest identification. No dashboard, no billing, no landing page. Point it at 500 Shopify/BigCommerce/WooCommerce stores pulled from a BuiltWith export, weighted toward Cloudflare-fronted domains.
• Day 3: Count. What percentage return a 403, a challenge, or a JS stub to GPTBot/ClaudeBot/PerplexityBot? This number decides everything. Below 10% and the pipeline was wrong about the ~27% claim — stop here.
• Day 4–5: Email the confirmed-broken merchants. Not a pitch — the finding, the timestamp, the URL, and the specific setting that caused it. Split the list: half get offered a $149 one-time fix report, half get offered a $39/mo monitor. Same finding, different ask.
• Day 5, decision: Go if ≥10% of probed sites are blocking AND ≥5 merchants pay for the monitor (not the one-time report) within 72 hours of the email. Both conditions, not either.

Falsifiable, and it fails fast: if merchants only want the one-time fix, I’ve learned in five days that this is a $149 service business — and I’d rather know that in week one than after building a monitoring SaaS nobody renews.