Agentic Commerce: How I Charge $2,200 to Make Stores Buyable by AI Agents

Something changed in my ecommerce clients' analytics this year, and it took me two months to name it. Sessions were flat. Conversion rate was up. And a growing slice of orders arrived from referrers I'd never had to think about before: assistants and shopping agents that read a product page, compared it against three competitors, and completed the purchase on the customer's behalf. Nobody browsed. Something machine-shaped did the browsing.

So I built an offer around it. I charge $2,200 to make a store readable, comparable and buyable by AI shopping agents, plus $450 a month to keep it that way. This is the whole playbook: what the audit actually checks, the fixes that moved numbers, what I charge and why, and the parts I still can't promise a client.

What "agentic commerce" means on a real store

Strip the buzzword and there are only two behaviours to design for.

The first is an agent that researches. A shopper types something like "quiet upright vacuum under $300 that works on wool rugs" into an assistant. The assistant fans out, reads product pages and reviews, and returns three options with reasons. Your store is either in that shortlist or invisible.

The second is an agent that transacts. The shopper approves an option and the agent proceeds: it picks a variant, checks stock, applies shipping to a country, and either completes checkout in a browser session or hands off through a merchant feed or checkout protocol.

Those two behaviours fail for opposite reasons. Research fails when your page doesn't state facts a machine can lift — the vacuum's noise rating in decibels, the actual weight, whether it's rated for high-pile carpet. Transaction fails when your funnel needs a human to guess: variant pickers that don't expose which option is out of stock, shipping costs revealed only after a login, popups that swallow the add-to-cart button, a checkout that hard-fails on autofill.

My audit splits along that seam. Half of it is about being chosen. Half is about being completable.

The four failures I find on almost every store

I've now run this on 11 stores, mostly $30k to $400k a month in revenue, across home goods, supplements, apparel and a B2B parts catalogue. The same four things come up.

1. Attributes live in pictures, not in text

The single most common problem. The size chart is a JPEG. The ingredient list is a JPEG. Compatibility ("fits 2018-2023 models") is baked into a lifestyle image. A human squints and understands; a research agent reads the page and finds nothing to compare, so it recommends the competitor whose page says "45 dB, 4.1 kg, safe for wool" in plain text.

Fix: every decision-relevant attribute appears as text, in a table, on the page. Not in a PDF, not in a tab that loads on click, not in an image's alt text alone.

2. Structured data that describes nothing useful

Most stores have Product schema because their theme shipped with it. It carries name, image, price and nothing else. No sku, no gtin, no brand, no additionalProperty entries for the specs, no shippingDetails, no hasMerchantReturnPolicy, and — my favourite recurring bug — availability hardcoded to InStock on every variant, including the four that have been sold out since March.

Fix: complete the offer object, per variant, generated from the same source of truth as the storefront so it can't drift. Wrong structured data is worse than none; an agent that quotes your stale price to a customer creates a support ticket and a trust problem.

3. Nothing corroborates the claim off-site

Agents cross-check. If your page is the only place in the world that says a product exists at that price with that spec, a cautious model hedges or drops you. Stores that get picked have their facts echoed in a merchant feed, on marketplace listings, in retailer spec sheets, in reviews that mention the specific attribute.

Fix: make the product feed accurate and complete first, then get at least two independent surfaces repeating the same numbers.

4. The checkout assumes hands and eyes

I test this by hand, slowly, the way a script would: keyboard only, no popups dismissed, autofill everything, one country changed halfway through. Typical breakages I've logged — a cookie banner that traps focus, a variant dropdown that only updates price via a hover event, "free shipping" that turns into $18 at step three, an email field that rejects the + character, and a coupon box that clears the cart when submitted empty.

Fix: every one of those is a normal accessibility and robustness bug. Making a store agent-friendly is mostly making it boringly correct.

The audit, step by step

It's five passes over about six working days. I do it in this order because each pass feeds the next.

Pass 1 — Prompt set (half a day). I write 25 buying prompts a real customer would type, spread across the client's top categories, price bands and objections. Then I run each one across three assistants and record what came back: which brands appeared, which pages got cited, what reasons the model gave. This is the baseline, and it's the part clients react to hardest — watching a competitor get recommended by name in eight out of ten prompts is more persuasive than any slide I could make.

Pass 2 — Page teardown (two days). Top 20 products plus the 5 highest-traffic collection pages. For each, I fill a sheet: attributes present as text, attributes present only in images, schema fields present and correct, price and stock accuracy against the admin, review count, and whether the page answers the top three objections from support tickets.

Pass 3 — Feed and off-site check (one day). Merchant feed completeness and error log, marketplace listings, any spec sheets a distributor publishes. I flag every place a number disagrees with the store. Disagreements are the thing agents punish.

Pass 4 — Transaction walk (one day). The slow, dumb, keyboard-only checkout run described above, on desktop and mobile, for three variants and two countries. I record video. Video ends arguments about whether a bug is real.

Pass 5 — Roadmap (one day). Everything found, sorted into three buckets: fixes that change what a machine can read about the product, fixes that change whether a purchase can complete, and content that needs writing. Each item gets an owner, an estimate, and a plain sentence explaining what it unlocks. That document is the deliverable.

What I charge, and what it costs me to deliver

The rate card, unchanged since April:

Offer	Price	What's included
Agent readiness audit	$2,200	5 passes, prompt baseline, teardown sheet, checkout video, roadmap, 60-minute walkthrough
Implementation sprint	$1,600 per two weeks	I do the schema, feed and page-template work with the client's dev
Monitoring retainer	$450 per month	Prompt set re-run monthly, feed error watch, stock and price drift alerts, one-page report

Direct costs on a $2,200 audit run around $90: assistant and API subscriptions, a screen-recording tool, and a feed-validation service. The real cost is my time — roughly 32 hours on the first audit in a category, 20 to 22 once I have a template for that vertical.

I price the audit as a fixed fee because scope is genuinely fixed: 20 products, 5 collections, 25 prompts, 2 countries. Anything beyond that is a second audit. The retainer exists because this decays — feeds break, products get renamed, themes get updated, and models change how they weight sources.

Numbers from the first four engagements

I only report what a client's own analytics showed, and I say plainly that none of this is a controlled experiment.

• Home goods, $120k/month. Prompt baseline: cited in 2 of 25. After schema rebuild, spec tables on 60 products and a fixed feed: 11 of 25 at 90 days. Assistant-referred sessions went from a rounding error to 3.4% of revenue.
• Supplements, $310k/month. Biggest win was ingredient tables replacing label photos. Cited 5 → 14 of 25. They also stopped 40% of a specific support-ticket category, which they cared about more than the citations.
• Apparel, $45k/month. Weakest result. Fashion prompts are taste-driven, not spec-driven, and the shortlists were dominated by marketplaces. Cited 1 → 3. I refunded the retainer after month two and told them to spend the money on paid social instead.
• B2B parts, $400k/month. Best result. Compatibility data was the whole product, and once it was structured per SKU, agents could answer "does this fit X" directly. Quote requests up 22% quarter over quarter; they attribute maybe half of that to the work.

Where this offer does not work

Say this out loud before taking money.

Taste-led categories. If nobody buys your product on specifications, structured specs won't get you shortlisted. Apparel, decor and gifting are hard.

Stores with no data source. If attributes only exist in a designer's PSDs, someone has to create them. That's a data project, not an audit, and it should be quoted separately.

Platforms you can't touch. Some hosted carts won't let you emit per-variant schema or fix focus traps in their checkout. Find out in the sales call, not in week two.

Clients who want a guarantee. I can't promise placement in an assistant's shortlist any more than I could promise a search ranking. What I promise is that the store will be accurate, complete, machine-readable and completable — and that I'll show them, with the same 25 prompts, what changed.

The stack I actually use

Nothing exotic, and I keep it cheap because the margin lives in hours saved, not tools.

For the prompt baseline I run the same 25 prompts by hand across three assistants and paste results into a spreadsheet with one row per prompt and one column per assistant. I tried automating it with APIs and stopped: the API answers and the consumer product answers differ enough that automating gave me a cleaner number and a worse picture of reality. Hand-running 75 prompts takes about three hours.

For the teardown I use a single sheet template with 18 columns per product, plus a schema validator and the platform's own feed diagnostics. For the checkout walk, a screen recorder and a keyboard. For monitoring, a scheduled job that pulls the feed daily and emails me when the error count changes or when a price in the feed disagrees with the storefront by more than a cent.

Contract terms that saved me twice

Three clauses I added after getting burned.

First, a no-placement-guarantee clause in plain language, initialled separately. It states that I optimise readability, accuracy and completability, and that no vendor controls whether an assistant recommends a product. Two prospects walked over this. Both were prospects I would have lost money on.

Second, a data-readiness condition: if attributes for the 20 audited products do not exist anywhere in the client's systems, the engagement converts to a data-creation project at an hourly rate, with a written go/no-go before that starts.

Third, a platform-access condition: I need the ability to edit templates and emit custom structured data. If the platform blocks it, the audit still ships but implementation is out of scope and the fee drops by $400.

How I get clients for this

Three channels, in order of how well they work.

The cold audit teaser. I run 10 prompts against a prospect's category, screenshot where a competitor gets recommended instead of them, and send it with two sentences. No deck. Roughly one reply in six, and about a third of replies book a call. It works because the loss is concrete and the fix is not obvious.

Existing SEO and CRO clients. Anyone already paying me for visibility work is an easy upsell — same goal, different reader.

Platform partners. Two Shopify agencies now hand me this piece because it sits awkwardly between their dev and marketing teams. Referral fee is 15% of the audit.

The 30-day version if you want to start

If you're a store owner and don't want to hire anyone: pick your ten best sellers. For each, write a plain-text spec table containing every number a buyer asks about. Fix your Product schema so price, stock and variants are generated from live data. Validate your merchant feed until zero errors. Then run ten buying prompts across two assistants and write down what you see, monthly. That sequence is 80% of what I get paid for, and you can do it in a month of evenings.

If you'd rather sell this as a service, do that same 30-day pass on one friendly store for free, document the before-and-after with screenshots, and use it as your only case study. That's how I got the first $2,200.