Use caseWorkflow automationDocument AI

Product data AI shopping assistants can read, filled from supplier specs instead of guesses

An honest take on agentic commerce: fill product attributes from supplier specs, publish clean feeds and schema.org markup, and track what AI assistants say.

A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, September 11, 2026.

The short version

A workflow that makes a product catalog readable by AI shopping assistants and Google's AI features. It audits which attributes are missing, fills them only from source documents such as supplier datasheets and manuals, normalizes units and variants, and publishes clean data to Shopify, Google Merchant Center and schema.org markup. People approve every claim about compatibility, safety or the environment. A monitor asks the assistants about your products each week and records what they get right and wrong.

Best for: Shopify brands with hundreds of technical SKUs whose product pages describe feelings rather than specifications.
Connects to: Shopify, Google Merchant Center, PIM (Akeneo or Plytix), schema.org Product markup, Google Sheets, Supplier documents
The AI does: Reads supplier datasheets and manuals, proposes attributes with the quote and page they came from, and asks assistants about your products every week.
People do: Approve claims about compatibility, safety and the environment, settle conflicts between documents, and decide which gaps are worth a supplier request.
Built as: AI Workflow Automation, usually $3.5K - $12K

What actually changed in AI shopping in 2026

Start with what did not work. ChatGPT's Instant Checkout launched in September 2025 on the Agentic Commerce Protocol, built with Stripe. By March 2026 OpenAI had sidelined it and moved checkout into apps that retailers such as Target run themselves, and Shopify's president pointed to the complexity of tax, inventory, shipping and subscriptions, as Digital Commerce 360 reported. Buying inside the chat window is not where the volume is.

Discovery is another matter. Shopify said AI-driven traffic and AI-driven orders to its merchants each tripled year on year in the second quarter of 2026, by its own count. Google launched the Universal Commerce Protocol with Shopify in January 2026, and Shopify says agent access to its catalog now takes just an API key. Assistants answer 'which 100W charger can run a laptop and a phone at once?' by reading structured product data.

Picture an electronics accessories brand with about 900 SKUs: chargers, cables, power banks, adapters, cases. The product pages were written for people scrolling: 'blazing fast charging for all your devices'. They do not say the output per port, whether a cable is rated for 5A, which charging protocols it supports, or which phones a case fits. An assistant cannot recommend what it cannot read, so it recommends a competitor whose data is complete.

None of this is about a flood of robot shoppers: Mastercard expects about one online shopper in ten to use agents to buy routinely by 2030. It is about being described correctly where people now ask. EU brands have one more reason, since ChatGPT ads were not sold in the EU as of September 2026 and accurate data is the only way into those answers.

The attributes assistants look for, and where each may come from

The audit compares every SKU with a target attribute set for its category. In this catalog most gaps are technical facts that already exist in a supplier document somewhere.

Attribute	Why assistants need it	Allowed source
GTIN, MPN and brand	Matching the product across Google, marketplaces and reviews	Supplier master data; never generated
Output per port and combined, in W	Answering 'can it charge my laptop and phone at once?'	Datasheet or test report
Charging protocols (USB PD, PPS)	Whether a phone or laptop fast-charges on it	Datasheet only; empty if not stated
Connector types and cable length	Filtering and side-by-side comparison	Datasheet, checked against photos for hero SKUs
Cable current rating (3A or 5A)	Whether a cable carries 60W, 100W or 240W	Datasheet only; empty if not stated
Compatible devices	The most common question, and the riskiest claim	Supplier compatibility list or own testing; always reviewed
Certifications (CE, UKCA, USB-IF)	Trust signals and marketplace requirements	Certificate or declaration of conformity on file
Manufacturer and EU responsible person	Required on EU listings under the GPSR	Company records
Shipping times and return window	Assistants compare delivery and returns across stores	Shopify and Merchant Center settings, published as structured data

Every value keeps a pointer to the document and page it came from. An empty field is allowed; an invented one is not.

The enrichment run for one SKU

The model reads documents and proposes values. It never fills a field from what it happens to know about similar products, and nothing it proposes reaches a live listing without passing code checks and, for claims, a person.

01 Trigger · Shopify Admin API, document folder
A SKU is new, changed or flagged by the audit
The nightly audit compares each SKU with its category's target attributes and lists the gaps. A new supplier document in a SKU's folder starts a run too.
02 Plain code
Collect the allowed sources
Only documents attached to that SKU count: datasheet, manual, test report, declaration of conformity. A SKU with no documents goes on the supplier request list instead of to the model.
03 AI model · Structured outputs
Extract attributes with citations
The model returns each attribute with the exact quote and page it came from, under a strict schema. 'Not stated' is a valid answer, and the preferred one when in doubt.
04 Plain code
Normalize and verify
Units converted to one convention (W, m, mAh, g), values checked against plausible ranges for the category, variants compared, and every quote searched for in the document text. A quote that is not there rejects the value.
05 Decision
Publish, review or reject?
Decided by attribute type and the checks, never by the model's confidence. At launch every type goes to review; a type moves to automatic publishing only after reviewers have accepted it unedited across a few hundred SKUs.
- A measurement with a verified quote: length, weight, connector, output then published automatically, once that type has earned it
- A claim about compatibility, certification, safety or the environment, or two documents that disagree then queued for a person with the quotes side by side
- No quote found, or a value out of range then rejected and logged; the field stays empty
06 Person · Google Sheets
A merchandiser works the review queue
Each row shows the proposed value, the quote and a link to the page of the PDF. Accept, edit or reject is one click, and rejected values are added to the extraction prompt as examples of what not to do.
07 System · Shopify, Merchant Center
Write once, publish everywhere
Accepted values go to Shopify metafields or the PIM, and the Merchant Center feed and the page's JSON-LD are generated from that one record, so the three cannot disagree. Large updates run as Shopify bulk operations within the API limits, and every write carries the record's version, so a retry never overwrites a newer edit.
08 Result
Keep the evidence
Every field stores its source document, page, reviewer and date. When a customer or a market surveillance authority questions a claim, the answer is one lookup away.

One datasheet in, attributes out, checks applied

A 100W two-port charger with a two-page supplier datasheet and a photo of the retail box. The model proposes; the checks decide what happens to each value.

Enrichment run, SKU CHG-100-2C

extract_attributes(sku: "CHG-100-2C", sources: ["datasheet_v3.pdf", "retail_box.jpg"])10 values proposed / 2 marked not stated: recycled content, cable included
verify_quotes(sku: "CHG-100-2C")10 of 10 quotes found in the source text / page references match
normalize(attribute: "output")"100W max (single port), 65W + 30W (both ports)" to max_output_w: 100, combined_output_w: 95
check_ranges(category: "wall_chargers")all values in range / combined output below the single-port maximum, consistent
classify(sku: "CHG-100-2C")7 measurements: publish / 2 claims for review: compatible devices (14 models), CE marking / 1 conflict: 182 g on the datasheet, 190 g on the box
The reviewer accepts the device list after checking the supplier's test report, confirms CE against the declaration of conformity on file, and takes 190 g from the newer box.
publish(sku: "CHG-100-2C")Shopify metafields updated / Merchant Center item updated / JSON-LD regenerated / change log written

Every value arrived with a quote that code found in the source. A person decided everything that is a claim rather than a measurement.

The same charger, written for scrolling and written to be read

Before

Title: 'Pro Fast Charger, Ultra Fast Charging, Black'
Description: 'Blazing fast charging for all your devices, perfect for travel'
No output per port, charging protocol or plug type in any structured field
Compatibility mentioned only in customer reviews, which contradict each other
Return window explained on a policy page an assistant has to find and interpret

After

Title with what people filter on: '100W USB-C charger, 2 ports, GaN, EU plug'
The same readable description, now above a specification table built from the same data
Output per port and combined, USB PD 3.0 and PPS, plug type and weight in metafields and the feed
A reviewed list of compatible devices, with the supplier's test report on file
Return window and shipping times published as structured data in Merchant Center and the page markup

Ways enrichment can make a catalog worse

The model fills a gap from memory

Ask a model for a charger's output and it will answer, plausibly and sometimes wrongly, from similar products it has seen. The pipeline is built around that failure: only attached documents count as sources, every value carries a quote that code finds in the document, and 'not stated' is a normal result. An empty field costs a recommendation. A wrong one costs a return, and possibly a complaint.

Availability that is out of date

An assistant that recommends an item sold out since Tuesday sends the shopper to a dead end, and Merchant Center disapproves items whose feed and landing page disagree on price or availability. Stock and price come from Shopify on every feed update, never from the enrichment store, and the feed diagnostics are checked daily.

Variants that blur together

A 1m and a 2m cable, or a case for two phone sizes, get their attributes mixed up when one datasheet covers a whole product family. The variant check flags any pair that differs in more than the attribute that defines the variant, and the reviewer sees both side by side.

Claims that need proof

'Works with every laptop', 'eco-friendly' and 'certified' are claims, not specifications. The EU's rules against unsubstantiated generic environmental claims (Directive 2024/825) apply from 27 September 2026, and the GPSR already requires safety information on EU listings. Environmental and safety wording goes to a person with the evidence; the pipeline never writes it.

Pages rewritten for bots

Attribute-stuffed titles and pages turned into specification dumps read worse for the people who still buy on them. Facts go into structured fields, feeds and markup, titles carry the few attributes people filter on, and descriptions stay readable. A change that only helps a crawler does not ship.

Reading too much into one answer

Assistants answer differently from day to day and from user to user. The monitor counts how often a statement appears across many runs of the same question instead of trusting a single reply. Search Console has its own quirks with machine traffic, covered in 3,572 impressions, zero clicks.

A weekly check on what assistants say about you

Most brands skip this part, and without it there is no way to know whether the work changed what assistants say.

Keep a list of about 60 questions real customers ask, from 'best 100W USB-C charger for travel' to 'will this charger run my laptop', in each market's language.
Run each question several times a week through the APIs of OpenAI, Google Gemini and Perplexity with web search switched on, and check a sample in the consumer apps each month, since API answers are close to, but not the same as, what shoppers see.
Record whether your products appear, which ones, the attributes each answer states, the URLs it cites and the competitors it names.
Compare the stated attributes with your catalog. A wrong wattage or a missing product becomes a finding, together with the source the assistant cited.
Send a one-page weekly report: new appearances, errors to fix at their source, and which catalog changes came before them.

What Shopify, feed tools and PIMs already cover

Start with what you have. Shopify syncs products to Google Merchant Center through its Google & YouTube app, supports category metafields from its standard product taxonomy, and exposes store catalogs to AI agents. Feed tools such as Feedonomics and Productsup transform and route feeds to many channels, and a PIM such as Akeneo or Plytix keeps master data in one place. If your supplier data is already structured and complete, these are enough, and you do not need me.

Custom work pays where those tools stop. They move and map data that exists; they do not read a supplier's PDF datasheet and fill in the output per port with a citation, and they do not tell you what ChatGPT said about your charger last week. For a catalog like this one, enrichment from messy supplier documents and the monitoring are the missing parts.

The usual shape is a hybrid: the pipeline writes into Shopify or the PIM, the feed tool keeps doing distribution, and an AI visibility tracker such as Profound can replace the self-built monitor if you want dashboards more than a report tied to your catalog. For listings across several marketplaces and languages, marketplace listing localization builds on the same clean record.

How you would know it is working

A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.

Attribute coverage per category: Share of SKUs with every target attribute filled from a cited source, tracked weekly. Values without a citation do not count.
Review acceptance rate: Share of proposed claims accepted without edits, per attribute type. A low rate points at a prompt or a supplier document problem, and decides which types may publish automatically.
Merchant Center item issues: Disapprovals and warnings per week from the Merchant Center diagnostics, split by cause.
Assistant mentions and accuracy: How often your products appear for the monitored questions, and how often the attributes stated match the catalog.
AI-referred sessions and orders: Sessions and orders with ChatGPT, Perplexity, Gemini or Copilot as the referrer, tagged in analytics, because this is the traffic the work is for.

What a build like this costs

This is built as AI Workflow Automation, which runs $3.5K - $60K overall. A build like this one usually lands in the single-step flow tier: $3.5K - $12K, 1-2 weeks. The first working version runs on your real data well before the end of that window.

What it costs to run

Extraction is a one-off cost per SKU plus re-runs when documents change; for a catalog this size, the model spend of a full pass is small next to the reviewers' time. The monitor's cost scales with questions times assistants times runs per week, and stays modest.

What moves the price

Catalog size, and how many categories need their own target attribute set
The state of supplier documents: clean PDFs, scans that need OCR, or nothing on file
How many claim types need review, which sets the reviewers' workload and the review screen
Where the record lives: Shopify alone, or a PIM with its own API and workflow
How many assistants, markets and languages the monitor covers

Who this is for

Electronics, tools and parts brands whose buyers choose on specifications the product pages do not state
Shopify brands with hundreds of SKUs and supplier documents sitting in folders instead of the catalog
EU brands that cannot buy their way into ChatGPT answers and need their products described correctly instead
Merchandising teams fighting Merchant Center disapprovals and marketplace attribute requirements at the same time

Questions people ask about this

What is agentic commerce, and is it real in 2026?

Agentic commerce means AI assistants finding, comparing and sometimes buying products on a person's behalf. The discovery part is real and growing: Shopify reports AI-driven orders to its merchants tripling year on year. The checkout part stalled, with OpenAI moving away from Instant Checkout in March 2026. For most brands the useful work is product data that assistants can read correctly.

How do I optimize my products for ChatGPT shopping?

Give assistants facts they can read. Fill the attributes buyers compare on from real supplier documents, keep titles specific, publish a clean feed to Google Merchant Center and complete schema.org Product markup with offers, shipping and return policy, and keep availability current. Then ask the assistants about your products regularly and fix what they get wrong. No ranking trick replaces correct data.

Can AI fill in missing product attributes?

Yes, from documents, not from memory. A model can read supplier datasheets and manuals and return each attribute with the quote and page it came from under a strict schema, as in structured outputs, and code can check every quote exists. Asked to fill gaps without a source, a model produces plausible values that are sometimes wrong, which is how catalogs end up listing the wrong wattage.

Does Shopify already make my products visible to AI assistants?

Partly. Shopify syncs products to Google through its sales channel app and exposes store catalogs to AI agents, so the plumbing exists. What no platform can do is pass on attributes your catalog does not contain. If a charger's output per port lives only in a PDF, nothing reaches the assistant until someone puts it into the product data.

How much does product data enrichment cost?

It depends mostly on the state of your supplier documents and how many claims need review, not on the model. A few hundred SKUs with clean datasheets is a small workflow project; thousands of SKUs with scanned documents in several languages and a review team is a larger one. Model spend is minor next to the time people spend approving claims.

Sources