ChatGPT visited a customer's product page yesterday. They asked it for a leather laptop bag under $200, it browsed three stores, and recommended one. Their store wasn't the one.
This is happening every day on every DTC storefront I've looked at. The merchant has no idea — their analytics dashboard says they had 4,300 visitors and a conversion rate of 2.1%. What it doesn't say: 412 of those visitors weren't humans clicking around. They were agents — ChatGPT-User, PerplexityBot, ClaudeBot, Google's Gemini crawler — doing the research a human used to do, then influencing what that human bought.
I've spent the last four months building tooling to see this traffic. Some of what I've found is what you'd expect. Some of it is genuinely surprising. The pattern isn't subtle — and the gap between merchants who can see it and merchants who can't is going to widen fast.
Here's what I've learned, what I think it means, and what to do about it if you operate a $1–10M DTC store.
What I'm actually seeing
Across the stores I've sampled — small set so far, ~12 DTC merchants in the $1–10M GMV band — agent traffic is consistently between 6 and 11 percent of session-level visits. Apparel and home goods skew higher; commodity consumables skew lower. The variance per store is mostly explained by category, brand recall, and how much the store's product catalog gets surfaced in AI assistant answers.
That isn't a "5% next year" number. It's a "your top 100 product pages are being browsed by ChatGPT this week" number. And in twelve to eighteen months it's going to be 25–40% of traffic, because the underlying behaviour shift is already complete: people who got used to asking ChatGPT how to do something at work are now asking it what to buy.
Concretely, here's the kind of session pattern that shows up on day one of having visibility:
- 47% of agent sessions read only product pages. No category browse, no search. The agent goes straight to a SKU because the user already asked it about a specific product elsewhere.
- 22% performs a price-comparison sequence: your product → out → competitor's matching SKU → out → your product again, sometimes 3–4 cycles.
- 11% bounces on the price page. Strong signal of a comparison-shop loss to whoever was cheaper.
- ~3% triggers a cart event before disconnecting. These are the highest-leverage missed conversions in the data — the agent recommended you and the human came in cold but didn't finish.
Standard analytics — GA4, Shopify Reports, anything keyed to bounce rate and session duration — filters all of this out. The agents fail human-shaped behavioural heuristics, get lumped in with "bot noise," and disappear from the dashboard the merchant actually looks at. The traffic is there. It just isn't legible.
Why this is happening now
Three things lined up.
1. Agent traffic became real, fast. OpenAI publishes the user agent strings their crawlers use. So does Anthropic, Perplexity, Google, Common Crawl, ByteDance. You can read the OpenAI bots reference, the Perplexity crawler page, the Anthropic robots.txt guidance — the documentation is right there. What's changed in 2025–2026 is that these crawlers stopped being curiosities and started being a measurable share of every storefront's traffic. Cloudflare's Radar dashboards show AI crawler activity climbing month over month with no inflection point in sight.
2. End-user behaviour caught up. People who asked ChatGPT to debug their code last year now ask it where to buy a sound machine. The shift from "search → click → buy" to "ask → buy" is partial — humans still browse, still click ads — but the agent is increasingly in the loop on the consideration step. By the time the human shows up on your storefront, the agent has already shaped what they're looking for.
3. The infrastructure to act on it is barely there. Most analytics platforms still treat agent traffic as a filtering problem. Most ad platforms can't bid on it. Most merchants don't know what their best-performing product for AI assistants even is. There's a window of 12–24 months where the operators who pay attention will compound an advantage on the operators who don't, and then the gap closes — partly because tooling catches up, partly because the agent-traffic share gets too big to ignore.
The combination of those three is why this is the year to start measuring. Not 2027.
The bet I'm making
I'm building Tactical on a deliberate two-step bet.
Step one (now): intelligence. Show merchants what agents are doing on their store. Which agents, what they look at, what they buy, where they leave. The wedge is informational — you can't act on traffic you can't see. The moat is time-decaying: as standards like ACP and AP2 mature, signed-agent traffic will make detection trivial and the measurement business will compress. That's fine — measurement isn't the endgame.
Step two (12–24 months): response. The durable product is merchant-deployable agents that respond to detected agent traffic. Not "your store has a chatbot." More like: an agent visited your high-intent SKU at 2am, your inventory is low, our agent kicks off a small workflow — pings you on Slack, holds inventory, surfaces a recommendation in the customer's email when they wake up. The intelligence wedge is the on-ramp. The agent layer is the destination.
Why DTC mid-market specifically — why $1–10M GMV stores and not the long tail or the enterprise? Because that's the band where:
- Agent traffic is already a measurable share, not a rounding error
- Operators have budget to actually act on it ($99–$500/mo SaaS works without procurement)
- The pace of decisions is fast enough to validate ideas in weeks, not quarters
- Trust in template-driven autonomy is realistic — the long tail is too price-sensitive, the enterprise too slow
What I'm deliberately not building:
- Marketplace integrations (Amazon, eBay, Walmart) — different product, post-purchase signal only
- Protocol-native runtimes for ACP/AP2 — standards still firming, that's a Phase 6 problem
- Visual no-code rule builders — premature; the work right now is figuring out what rules even matter
- Enterprise — different distribution motion, doesn't fit a solo founder's pace
Holding scope is the hardest part. Most merchants ask for the next thing five minutes after seeing the first thing. The discipline is shipping the wedge fully before chasing the agent layer.
The agent-ready DTC playbook
If you operate a DTC store and you've read this far, here are the five things to actually do. Each is DIY-able with a half-day of work; if you'd rather skip the DIY, Tactical does some of it.
1. Stop filtering agents as bot noise
The first move is just turning on visibility. In Google Analytics, "bot traffic" is a single bucket — useful for blocking scrapers, useless for distinguishing "ChatGPT visited my product page on behalf of a real customer" from "scraper trying to rip my catalog."
What to do:
- Look at your raw server logs (or your CDN logs — Cloudflare, Fastly, Vercel) for the past 30 days. Filter to user agents matching
ChatGPT-User,GPTBot,PerplexityBot,ClaudeBot,Google-Extended,Amazonbot,Bytespider. Count the sessions. - For each agent type, count how many distinct product URLs it touched.
- Sum the time-on-page across those sessions. You'll be surprised.
You're looking for a baseline. Number isn't the point — establishing that this traffic exists and is non-trivial is the point. If you're at less than 1% of total traffic, you can probably wait. If you're at 5%+, you have a hole in your analytics.
2. Classify agents apart from each other
The second move: not every agent is the same. ChatGPT-User (the agent that browses on behalf of a logged-in ChatGPT user asking it a question) is fundamentally different from GPTBot (the agent that crawls for training data). Both come from OpenAI; only one is a sales signal.
What to do:
- Tag agent sessions by intent class.
ChatGPT-User,PerplexityBot,ClaudeBotwhen invoked as "Claude for Chrome" — these are session-level shoppers. GPTBot,Google-Extended,CCBot— these are catalog scrapers. Important for SEO, less actionable for sales.- Apply this tag at session ingest. Most off-the-shelf analytics won't do this for you. You'll need either a small classifier (regex over UA + behavioural heuristics) or something purpose-built. Tactical does this; so does any half-decent CDN with bot management if you write the rules yourself.
The output you want: a per-shop breakdown of "shopping agents" (the ones whose presence correlates with a real customer asking for advice) vs "training/SEO crawlers" (the ones whose presence is about long-term indexing).
3. Make your structured data agent-friendly
Agents read your storefront the way a screen reader does — they need the facts about a product, not the marketing copy. The single highest-leverage thing you can do for agent visibility is to make sure your product pages emit clean, complete Product schema in JSON-LD.
What to do:
- Open one of your top product pages. View source. Look for
<script type="application/ld+json">. Validate the contents at validator.schema.org and Google's Rich Results Test. - Ensure every product surfaces:
name,description,image,offers(withprice,priceCurrency,availability),brand,sku, and ideallyaggregateRatingandreviewif you have them. - For variant-heavy products: emit a separate offer per variant or a
hasVariantarray. Agents handle this fine; they don't handle a flat "starts at $X" string. - If you're on Shopify, the standard themes get most of this right but often skip
aggregateRating. Worth checking. - If you're on WooCommerce, the default schema output is genuinely poor — install one of the SEO plugins that emits proper Product schema (Rank Math, Yoast Local) or write the schema yourself.
This is the work that compounds. Better structured data → better agent recommendations → more agent-driven sessions → more revenue. The agents reward stores that make their job easy.
4. Get a human in the loop for high-intent agent sessions
When an agent on your store is doing the slow, deliberate version of a buying decision — viewing a product, reading specs, checking inventory, checking the price, going away, coming back — you want to know in real time. Not the next morning. Not in the weekly report.
What to do:
- Define what "high-intent agent session" means for your store. A working baseline: agent visitor + 3+ product pages + cart event + 2+ minutes session length.
- When that pattern fires, post to your Slack — channel name, agent type, product, country, dashboard link. Don't sit on it for a day; the agent's recommendation is being made now.
- Have a written response: most merchants haven't thought about what they'd actually do when they see a high-intent agent session, so the alert just becomes noise. Decide first. Maybe it's an inventory hold, maybe a pricing tweak, maybe nothing — but the answer should be deliberate.
Tactical ships this as the High-Intent Slack Bridge template. You can also build it yourself with any analytics tool that supports webhook outputs and a 20-line Slack incoming-webhook script.
5. Welcome the agents you want
This is the most contrarian recommendation. Most operators reading this far are thinking about agents as a threat — scrapers, bot traffic, lost revenue. That framing misses something: a meaningful share of agent traffic to your store is trying to recommend you. They want to give your product as the answer. They just need you to make it easy.
What to do:
- For agent visitors specifically, surface a small banner pointing them at your structured product feed, your spec sheets, or your
llms.txtif you have one. Something machine-friendly that says "we're agent-friendly, here's the data you want." - Add an
llms.txtfile to your root domain — there's no formal spec yet, but several major LLM crawlers respect it. Anthropic and Perplexity have both shipped support; OpenAI is rumored. Worth doing pre-emptively. - For your top-20 products, write a one-paragraph plain-text spec dump that strips marketing copy. Make it accessible at
/products/{slug}/specor similar. Agents will use it.
Tactical ships an Agent Welcome banner template. Or you can hand-roll a <script> tag that does the same — match a UA pattern, inject HTML, ten lines of code.
What's next
The five-step playbook above is roughly where the intelligence wedge ends. It's enough to give you visibility, classification, basic response, and a friendlier surface for agents that want to recommend you. Doing all five takes a serious DTC operator a couple of weeks.
The next layer — what I'm building toward over the next 12–24 months — is where the agent layer becomes the product. Stores deploying their own small agents that respond to detected agent traffic, customize answers per visitor, hold inventory deliberately, route comms intelligently. Right now that's a per-merchant consultative engagement, not a self-serve tier. We're working with a handful of stores at a time on what their bespoke setup actually looks like, and pulling general-purpose features out only after we've seen the same need from two or three merchants.
If that's interesting — if you're a DTC operator in the $1–10M band who's curious what your agent traffic looks like, or you've already done the playbook and you want help with the agent-layer setup — I'm running these consultatively right now. The fastest path is the Enquire form on the Tactical billing page, or just reply to this post. Either way, I read everything personally.
If the playbook is enough for you, Tactical's free tier gives you 100 agent sessions per week to validate steps 1 and 2 against your actual store, and the paid tiers cover steps 3–5 if you'd rather not DIY. No sales motion — sign up, paste a snippet, see your data within five minutes.
The window is open right now. Twelve months from now, agent traffic on DTC storefronts will be impossible to ignore — and the operators who started measuring it in 2026 will have a year of pattern data on every other operator who didn't. That's the bet I'm taking.
Sourabh — building Tactical, solo. Found this useful, or disagree with the bet? Email me at [email protected] — I read everything.