Stickybit.PortuguêsE-commerce and AI agents · 2026
E-commerce and AI agents

Your next customer has no eyes.

ChatGPT, Gemini and Meta's Muse already choose where to buy on the customer's behalf. They see no banner and no storefront: they read the HTML, the product data and whatever your firewall lets through. See your store the way they do.

Specimen · the same page, for a customer and for an agent
← Storefront: what the customer seesReading: what the agent receives →

Example store. Every failure on the right showed up in at least one of the four large retailers we read in September 2026.

Day to day

It is a customer who calls instead of walking in.

Think of a customer who does not visit the store: they phone. They ask the voltage, the installed price and whether it arrives by Friday. If the clerk cannot answer, they call the next store. They never saw the window, the banner or the promotion.

The AI assistant that shops for the customer (ChatGPT, Gemini, Meta's Muse) is that customer on the phone. It does not run the page like a browser: it downloads the HTML, looks for the spec sheet written for machines (structured data) and the product barcode, so it knows your air conditioner is the same one sold next door. If the price only appears after JavaScript runs, for the agent there is no price.

And there is the doorman. Many stores welcome AI bots on the sign on the door (the robots.txt) and then the firewall blocks those same bots on the third request. The store thinks it is open; the agent finds the door shut and leaves.

the customer asks: "12k BTU split, 220 V, under R$ 3k" the assistant reads the HTML, not the storefront Store 1 price: ✗ (JS only) voltage: ✗ barcode: ✗ Store 2 doorman: 403 could not read Store 3 ✓ chosen price: R$ 2,899 220 V · 12.000 BTU EAN 789…4 ✓
Illustrative: the assistant only compares what it can read. The store that delivers readable price, voltage and barcode is the one that makes it into the answer.
Measured

What we found at four large retailers.

Appliances, furniture and pharmacy. All with stores built well for people. We downloaded 22 to 24 product pages from each, without running JavaScript, the way AI bots do, and ran real shopping searches. For agents, it is a different story.

4 of 4without a barcode an agent could use to recognize the same product at another store
1 of 4hid the price behind JavaScript and a ZIP code. Asked "where to buy cheapest", the answer cited a competitor
1 of 4welcomed AI bots on the door sign and blocked them at the firewall
0 of 3showed up for "near me" in their own home city. The neighborhood store with a simple page did

Store A · home appliances

24 product pages · VTEX platform
Text the bot reads, without JavaScript1,897 chars
Spec sheet in the HTML0/24
Description in the machine sheet0/24
Valid barcode0/24 (internal code)
Readable review score0/24
AI bots get inyes
Shows up in shopping searches0/7
Items in order1 / 7
Brand search: 1st place. The problem is not reputation, it is being found.

Store B · appliances & furniture

22 product pages · server-rendered
Text the bot reads, without JavaScript6,394 chars
Spec sheet in the HTML18/22
Description in the machine sheet22/22
Barcodemissing in 22/22
Readable review score1/22
AI bots in a burstblocked (403)
Shows up in shopping searches1/11
Items in order3 / 7
The sign on the door (robots.txt) welcomes AI bots by name; the doorman (firewall) blocks them after the first 2 or 3 requests.

Store C · pharmacy

24 product pages · server-rendered
Question-and-answer text in the HTMLyes
Guide for AIs (llms.txt)well written
Machine sheet of any kind0/24
Price in the HTML0/24 (only after ZIP code)
Service pages cited3 of 7 searches
"Where to buy cheapest"cited a competitor
"Fast delivery" in its home cityabsent
Items in order3 / 7
It wrote for machines but did not structure: the agent knows what the medicine is, not what it costs.

Store D · pharmacy

24 product pages · VTEX platform
Price in the HTML24/24
Machine sheet with price and stock24/24
Valid barcode2/24
Readable review score0/24
Guide for AIs (llms.txt)returns the home page
"Where to buy cheapest"cited, with price
Local searches in its home cityabsent
Items in order3 / 7
The mirror of store C: the agent sees the price and wins the price question; it vanishes on "near me".

Pilot sample, September 2026: 22 to 24 product pages per store, sampled from the sitemaps and downloaded without JavaScript. The shopping searches are an approximation (one web search engine, one run per question); measuring ChatGPT, Gemini, Perplexity and Muse is the next phase. The firewall test ran from a single address. Stores anonymized.

Before it picks your store

What the agent checks: seven points.

In our reading, each score comes with the address, the date and the snippet that proves it. Here you can do a quick self-check: tick what you think your store already has.

Agent checklist

1. Access?
2. Reading?
3. Structure?
4. Price?
5. Place?
6. Identity?
7. Purchase?
Your store, by your count0 / 7
Tick the points on the left.
How it works

It starts free and only becomes a project if the reading shows it pays off.

24h
Free · within 24 business hoursFree readingYou give us the address and get by email what an agent can read from your store, with the proof for each item.
2 to 3 weeksFull diagnosisWe run 120 shopping questions on ChatGPT, Gemini, Perplexity and Muse and show where you appear, where competitors appear and why.
<Product>gtin: 789…
4 to 8 weeksFixesWe make the store readable for agents: spec sheet in the HTML, structured data, barcodes, regional prices, local pages, firewall.
MCP
MonthlyConnection and measurementWe connect the store to agent purchase protocols where the platform does not, and track every month what changed.
Track record

We have fixed e-commerce under pressure.

Two cases at one of Brazil's largest home-appliance e-commerces, and the next channel stores will have to serve.

2019A proxy lets 14 checkouts share one TOTVS license on the eve of Black Friday.
2020In-house shipping replaces a SaaS in 67 days, on R$ 150 a month of infrastructure.
2026We read four large retailers the way AI agents read.
2027The AI agent becomes a sales channel. Stores need to be ready for it (our reading).
Who does it

Software engineering in Porto Alegre since 2004.

The people you talk to are the people who write the code, from diagnosis to production.

Evidence in every score

No "could be better": every item comes with the page, the date, the response code and the snippet.

Go in production

The reading runs on our own code, with no third-party tool in the middle.

No impossible promises

We guarantee what can be verified and measure the rest every month.

Next

Agents that work for you.

With a clean catalog, the same data feeds agents that operate your store: they adjust prices within ranges you set, warn about stock-outs before they happen and log every decision.

Every action goes through an approval checkpoint you control. What is within range goes ahead; what is outside goes back to a person. How we govern agents in production.

store agent price −8% restock 40 units price −35% approval ✓ within range ✓ within range ✗ back to you log of every decision: who asked, what changed, when
Illustrative: ranges and approval are set by the store; the log keeps who asked, what changed and when.
FAQ

What people always ask.

Is this SEO under another name?

No. SEO competes for a position in a search engine. Here the goal is for an agent to be able to read, compare and trust your catalog. Most of the work is verifiable engineering: the HTML, the structured data, the firewall.

Do you guarantee I will show up in ChatGPT?

Nobody can honestly guarantee that. We guarantee what can be verified: that the agent can get in, read and understand your store, and we measure every month where you show up.

My platform already integrates with Gemini. Do I need this?

The integration connects the pipe. If the catalog flowing through it has no barcode, hides the price or keeps the spec sheet only in JavaScript, the agent gets little and picks another store.

What happens to my store's data?

The reading only touches public pages. Any deeper test, such as the firewall one, is only done with written authorization.

Free reading · within 24 business hours

See your store the way the agent does.

You get by email what an agent can read from your store, with the proof for each item. Public pages only; our bot identifies itself as Stickybit and respects robots.txt.

Three words on this page
Barcode (GTIN)

The product number that is the same in every store. It is how the agent knows two listings are the same air conditioner. A store's internal code in its place is worse than empty. More

Machine sheet

Structured data: name, price, stock, voltage and rating written in a format the agent reads directly, without guessing from text. More

The doorman

The site firewall. It can block AI bots even when the sign on the door (robots.txt) says they may come in. More

Limits

Where this could be wrong.

Small sample

22 to 24 pages per store, at four stores. Enough to find defects repeated on 100% of pages (like the barcode), not to grade an entire store.

Approximate search

Shopping searches ran on one web search engine, once each. Real ChatGPT, Gemini, Perplexity and Muse answers vary between runs; measuring that is the full diagnosis.

The doorman seen from one address

We tested the firewall with the bots' names, from our address. A firewall that checks the bot's real address would be right to block a fake "GPTBot"; only the store's logs confirm.

Nobody guarantees showing up

2026 studies show no technique has a stable effect on being discovered by AIs. What is solid: being readable, having data only you have, and measuring every month.

See also
Sources

← stickybit.com.br