Data type · Agent feeds

Product feeds an AI shopping agent can actually act on

Three companies asked us for this last quarter, all building natural-language product discovery. The feed they need is not the catalogue feed with a different label on it.

Feeds for AI shopping agents are product records built for machine consumption rather than a dashboard: typed and unit-normalised attributes so natural-language constraints resolve, freshness stamps an agent can act on, and — the part almost nobody gets right — absent attributes marked explicitly absent. An agent that cannot distinguish no from we don't know will recommend confidently and wrongly.

Three companies asked us for this last quarter, all building natural-language product discovery. The feed they need is not the catalogue feed with a different label on it.

Get a free agent-feed sample See sample data

Free pilot on your own sources, returned in 24 hours. No card, no trial clock — and you keep the sample data either way.

✓ Attributes typed and unit-normalised✓ Nulls carry a reason, always✓ Freshness stamped per record

agent_feed.jsonl LIVE FEED

// one record per product, built for a machine consumer {"product_key": "aw-4471", "title": "Example 1.7L stainless kettle", "attributes": { "capacity_ml": {"value": 1700, "basis": "parsed_spec_table"}, "weight_g": {"value": 1240, "basis": "parsed_spec_table"}, "dishwasher_safe": {"value": false, "basis": "stated_by_manufacturer"}, "cordless": {"value": null, "reason": "not_published"}}, "price": 44.99, "in_stock": true, "observed_at": "2026-08-25T06:00Z", "freshness_minutes": 184, "safe_to_recommend_until": "2026-08-25T18:00Z"} // the distinction that decides whether an agent is trustworthy: // dishwasher_safe = false -> the maker says no // cordless = null+reason -> nobody said either way // collapse those two and the agent asserts things nobody claimed

3 of 8,204,110 product records attribute fill reported per field · nulls reasoned · schema v1.0

Our Data Powers

B2C Marketplace

amazon

D2C + Marketplace

NYKAA

D2C + Marketplace

Walmart

FMCG Marketplace

udaan

Food Delivery

Uber Eats

Quick Commerce

blinkit

Taxi Aggregator

Uber

E-Commerce

Tmall

Summary What it is What's included Schema Markets Who it's for Examples How we engage SLA FAQ

Written and reviewed by the Actowiz Solutions Data Engineering team Published 5 August 2026 Last updated 5 August 2026 Reading time 10 min

Key facts at a glance

Consumer: A machine, not an analyst or a dashboard
The critical rule: Absent is not false. Null carries a reason, always
Attributes: Typed and unit-normalised, so range constraints resolve
Provenance: Every attribute carries how it was derived
Freshness: Stamped per record, with a stated safe-to-recommend window
Coverage: Per-attribute fill rate reported, so gaps are known before deployment
Not included: Embeddings, rankings or recommendations — those are yours
Refresh: Driven by what your agent promises, not by a default cadence

3companies asked for this last quarterall building natural-language discovery

100%of nulls carry a reasonno bare empties reach the agent

0inferred attributes in observation fieldsa guess is never presented as a fact

24hfree sample on your own cataloguewith per-attribute fill rates

Key takeaways

Consumer: A machine, not an analyst or a dashboard
The critical rule: Absent is not false. Null carries a reason, always
Attributes: Typed and unit-normalised, so range constraints resolve
Provenance: Every attribute carries how it was derived
Freshness: Stamped per record, with a stated safe-to-recommend window
Coverage: Per-attribute fill rate reported, so gaps are known before deployment

Last verified 5 August 2026 by the Actowiz Solutions Data Engineering team.

Definition

Why the catalogue feed does not work for an agent

A person reading a product page fills gaps automatically. They see no mention of a feature and conclude nothing — they check elsewhere, or they ask. An agent does not do that unless the data tells it to.

Three things that break

Marketing copy does not resolve constraints. "Compact and lightweight" cannot answer under 500 grams. The attribute has to be a typed number in a known unit.
A missing attribute reads as a negative. If dishwasher_safe is absent and the agent filters for dishwasher safe, the product is excluded — and the shopper never learns it might have qualified.
Stale availability becomes the merchant's problem. An agent recommending an out-of-stock item damages the merchant, not the service vendor. It needs to know how old the record is.

What we deliver instead

Every attribute as {value, basis} where it is known, and {value: null, reason} where it is not. The reason distinguishes not published, published but unparseable, and not applicable to this category — three situations an agent should handle differently.

Plus observed_at, freshness_minutes and a safe_to_recommend_until horizon agreed with you, so the agent can degrade gracefully rather than asserting a stale price.

What the feed carries

Six things an agent needs that a catalogue feed lacks

All of it derived from the same collection, shaped for a different consumer.

Typed, unit-normalised attributes

So numeric and range constraints resolve.

Values typed, not strings
Units normalised to a stated base
Ranges represented as min and max, not collapsed
Parse basis on every derived value

Explicit absence

The single most important field family here.

Null with a reason on every unknown attribute
not_published, unparseable and not_applicable distinguished
Never defaulted to false
Per-attribute fill rate reported per category

Freshness an agent can act on

So a stale recommendation is preventable.

Observation timestamp per record
Freshness in minutes at delivery
A safe-to-recommend horizon agreed with you
Availability and price refreshed independently of attributes

Stable identity

So an agent's memory of a product survives.

Persistent product key across runs
Cross-retailer links with match confidence
Variant axes explicit, not flattened
Delisting recorded as an event, not a disappearance

Offer context

Because the price alone is rarely the offer.

Price with currency and market
Delivery promise where collected
Promotional mechanics as structured fields
Fees kept separate from item price

Compliance-safe content

So an agent does not repeat a claim as fact.

Manufacturer claims marked as claims
verified_by_actowiz false on every claim field
No inferred attributes in observation fields
Source recorded per attribute

Service scope

What the ecommerce data scraping service includes

A managed engagement, not a tool licence. We own the pipeline and everything that breaks in it.

✓ Included in every engagement

Attributes typed and unit-normalised, each with its parse basis
Explicit nulls distinguishing not_published, unparseable and not_applicable
Per-attribute fill rate reported per category, in every batch
Freshness in minutes plus an agreed safe-to-recommend horizon
Persistent product keys and explicit variant axes
Cross-retailer links with match confidence
Manufacturer claims marked as claims, never as facts
Source discovery, scoping and a written collection plan
Free pilot on your own sources before any commitment
Full pipeline build, hosting and proxy infrastructure
Schema design, validation and sampled human QA on every run
Ongoing maintenance when source layouts change — our cost, not yours
Delivery to your warehouse, bucket, SFTP or API endpoint
Documented methodology and compliance notes for your legal review

× Not included — stated upfront

Embeddings, vectors, ranking or recommendation logic
Attributes inferred by a model and placed in observation fields
Natural-language summaries generated from the data
Machine translation written into observation fields
A blank attribute left indistinguishable from a stated negative
Anything behind a login, paywall or credentialed session
Personal data beyond a documented lawful basis
Licensed third-party datasets we do not hold rights to
Guarantees about fields a source simply does not publish

Schema

Agent feed fields

One record per product, shaped so a machine can act without guessing.

Agent feed schema v1.0 — abbreviated
Field	Type	What it captures	Refresh
`product_key`	string	Persistent across runs, so agent memory survives	Every record
`title / brand / category`	string	As published, plus normalised category	Every record
`attributes`	object	Each as {value, basis} or {value: null, reason}	Every record
`attribute_fill_rate`	object	Per-attribute fill for this category, so gaps are known	Per batch
`price / currency / market`	number / string	Offer basics with market as a dimension	Every record
`in_stock / availability_reason`	boolean / string	Availability with a reason where false	Every record
`observed_at / freshness_minutes`	timestamp / number	How old the record is at delivery	Every record
`safe_to_recommend_until`	timestamp	The horizon you agreed, so the agent can degrade gracefully	Every record
`variant_axes`	array	Explicit rather than flattened into the title	Where applicable
`match_confidence`	number	On cross-retailer links, so weak matches can be excluded	Matched records
`verified_by_actowiz`	constant	False on claim fields. A claim is never a fact	Claim fields

The per-attribute fill rate ships with every batch. An agent deployed on an attribute that populates on 30% of a category will fail in a way that looks like a model problem and is not.

Coverage

Where agent feeds are strongest and weakest

Attribute density decides this, and it varies enormously by category.

Consumer electronics — strong, dense spec tablesHome appliances — strongIT hardware — strongIndustrial and MRO — strong on part attributesKitchen and small appliances — goodFurniture — moderate, dimensions carry identityTools and hardware — goodAutomotive parts — strong on fitmentBeauty — moderate, shade adds an axisFashion — weaker, attributes are descriptiveGrocery — moderate, pack and nutrition parse wellToys — weaker, age guidance is inconsistentBooks and media — strong on identifiers, weak on attributesHandmade — not suitable, no stable productPerishables — availability moves faster than a horizon can holdServices — not a product feed

We report the per-attribute fill rate on your own categories in the sample. A category where your key constraint attribute populates thinly is one where an agent will disappoint users, and it is better to know that before launch. Request a source we don't list →

Markets served

Countries and markets where this service is in highest demand

We deliver into 40+ countries. These are the markets where this particular service is requested most, and the reason demand concentrates there.

Highest-demand markets for this service, and why demand concentrates there
Market	Why demand concentrates here

North America

United StatesCanadaMexico

United Kingdom & Ireland

United KingdomIreland

Western Europe

GermanyFranceNetherlandsBelgiumSpainItalySwitzerlandAustria

Nordics

SwedenNorwayDenmarkFinland

Middle East

United Arab EmiratesSaudi ArabiaQatarKuwaitIsrael

Asia Pacific

SingaporeAustraliaNew ZealandJapanSouth KoreaMalaysiaIndonesiaThailandVietnamPhilippines

South Asia

IndiaBangladeshSri LankaPakistan

LATAM

BrazilArgentinaChileColombia

Africa

An agent asserting an unverified manufacturer claim is a liability.

What we deliver

Claims marked as claims, no inferred attributes, source per attribute.

Metric that moves

Claim exposure

Use cases

How agent feeds get used

Four patterns, and one of them is defensive.

Natural-language constraint resolution

Typed and unit-normalised attributes so 'under 500 grams', 'fits a 27-inch monitor' and 'dishwasher safe' resolve against real values rather than being matched against marketing copy.

Outcome: Answers that hold up when a user checks them.

Preventing confident wrong recommendations

Explicit absence with a reason, so an agent can say it does not know instead of asserting a negative nobody published.

Outcome: The failure mode that damages trust fastest, removed.

Stale-recommendation prevention

Freshness in minutes plus an agreed horizon, so an agent degrades to 'check availability' rather than confidently recommending something out of stock.

Outcome: Merchant relationships that survive the agent.

Agent discoverability audit for brands

Attribute coverage on your own products against competitors, showing where your catalogue cannot answer the constraints buyers will ask agents about.

Outcome: A fixable content gap identified before it costs share.

Engagement examples

Two engagements, anonymised

Clients rarely permit naming. These are real engagement shapes with identifying detail removed, so you can judge whether the work resembles your situation.

Examples are anonymised at client request. Named references are available on request under NDA. See published case studies →

The 24-hour sample — run on your sources, not ours

Before you commit to anything, we run this service against your own sources and send you the output. If the coverage isn't there, the sample will show you that too — which is the point. We would rather lose the deal at the pilot than at month three.

Real extraction from your actual sources
Returned within 24 hours
Coverage and QA note included
You keep the data either way
No card, no trial clock
Named engineer on the call

Get my free sample Book a 20-min scoping call Reply within one business day. Reference calls available under NDA.

How we engage

Three ways to engage us

Same collection pipeline and QA underneath. The difference is who holds the schedule and how the data reaches you.

Managed service (most common)

We own the collection, the QA and the delivery. You receive clean data on a schedule and never touch a scraper.

Dedicated engineer assigned to your account
Site changes fixed by us, not reported to you
Scheduled delivery to your warehouse or S3
Named contact on Slack or email

Best fit: Teams who need the data, not the infrastructure.

API access

The same collection pipeline exposed as an authenticated REST endpoint your systems query directly.

On-demand and scheduled endpoints
Rate limits agreed to your load profile
Sandbox keys for integration testing
Versioned schema with deprecation notice

Best fit: Product and engineering teams building on live data.

One-time or project extraction

A defined pull for a specific question — market sizing, diligence, a pitch, a one-off audit.

Fixed scope agreed in writing upfront
Single delivery with full QA report
Methodology documented for your records
Converts to managed if you want continuity

Best fit: Research, strategy and diligence work with a deadline.

Pricing

Every engagement is quoted individually, because the honest answer depends on your scope: how many sources, how many records, how often, and how the data reaches you. We scope it with you, run a free pilot on your own sources, and then quote a fixed monthly figure — no per-request metering and no overage billing when volumes move. Request a quote and you will have a number after one call.

Build vs buy

Reshape your catalogue feed, or take an agent feed?

If your catalogue already has typed attributes with explicit nulls, reshape it. Most do not.

In-house build vs self-serve tool vs Actowiz managed service
Consideration	In-house scraping team	Generic proxy / DIY tool	Actowiz managed feed
Time to first usable data	6–12 weeks of engineering before anything is trustworthy	Days, but output needs manual cleanup before use	Free pilot in 24 hours, production in 5–10 business days
Who fixes it when a source changes	Your engineers, at the cost of their roadmap	You do — tools report failures, they don't resolve them	We do, same business day, inside the retainer
Data quality assurance	Whatever your team has time to build	None beyond HTTP success	Schema validation plus sampled human QA on every run
Compliance documentation	Rarely produced, then requested urgently by legal	Not provided; terms risk sits with you	Sources, method and lawful basis documented for review
Accountability	Distributed across a team with other priorities	A support ticket queue	A named engineer and an account owner
True annual cost	Engineer salaries, proxies, hosting, ongoing maintenance	Low licence fee plus significant hidden analyst time	One fixed monthly retainer, quoted after scoping

The distinction the whole feed is built around

There are two ways an attribute can be not-true, and an agent must handle them differently.

	The maker said no	Nobody said
Our field	`{"value": false, "basis": "stated_by_manufacturer"}`	`{"value": null, "reason": "not_published"}`
Agent should	Exclude confidently	Include with a caveat, or ask
If collapsed	The agent excludes products that might qualify, and tells the user they do not — asserting something nobody claimed

Almost every product feed in the market collapses these, because a dashboard user reading a blank cell knows what it means and an agent does not.

Three reasons are distinguished, not one: not_published (the source is silent), unparseable (it was stated but not in a form we could type confidently), and not_applicable (the attribute is meaningless for this category). An agent can reasonably surface the second to a user and should ignore the third.

What we deliberately leave to you

This is a data feed, not an agent, and the boundary is worth stating because the two are often conflated.

Not included

Embeddings or vectors. Yours to generate, against your model, on your schedule.
Ranking or recommendation logic. That is your product.
Inferred attributes. If you want to predict whether an item is dishwasher safe from its material, run that model — and keep the output in a column that is visibly a prediction.
Natural-language summaries. We deliver structured fields; generating prose from them is your layer.

Why we stop there

The moment we generate an attribute, your agent's provenance chain breaks. You could no longer tell a user — or a regulator, or a merchant — which facts came from the manufacturer and which came from a model.

That distinction is the thing agents will be judged on, and it is cheap to preserve now and expensive to reconstruct later.

How it works

How an agent feed goes live in 5 to 10 business days

The sample exists mainly to show you your real attribute coverage.

Tell us the constraints your agent must answer

Not the fields — the questions. 'Under 500g', 'fits a 27-inch monitor', 'suitable for sensitive skin'. Those determine which attributes matter.

We report per-attribute fill rates on your categories

Before quoting. A key constraint attribute that populates on 30% of a category is a launch problem, and you should see it now.

Free sample within 24 hours

Real records from your own catalogue with typed attributes, reasoned nulls and freshness stamps, so you can run your agent against it before committing.

Agree the safe-to-recommend horizon

How stale is too stale depends on your category and your promise. We set it per field group rather than one number for everything.

Production and monitoring

Live in 5 to 10 business days, with the per-attribute fill rate in every batch so coverage degradation is visible before users notice it.

Formats & destinations

JSON or JSONL for agent consumption, Parquet for bulk loads, or a REST endpoint with per-product lookup. Delivered to Amazon S3, Google Cloud Storage, Azure Blob, Snowflake, BigQuery or an endpoint you specify.

Compliance & data ethics

We collect only publicly accessible information. Manufacturer claims are captured as claims with verified_by_actowiz set false, and no attribute is inferred or modelled into an observation field.

Service commitments

What we commit to, in writing

These are contractual, not marketing copy. They appear in the engagement document.

Service level commitments written into every managed engagement
Commitment	What we hold ourselves to
Pilot turnaround	A real sample from your own sources within 24 hours of scoping, at no cost.
Go-live	Production collection running within 5–10 business days of sign-off.
Delivery punctuality	99.5% on-schedule delivery, measured monthly and reported to you.
Breakage response	Source layout changes triaged same business day; critical sources inside 4 hours.
Data quality	Schema validation on every run plus sampled human QA before any delivery leaves us.
Escalation	A named engineer and an account owner, not a shared ticket queue.
Change requests	Field additions and source changes handled inside the retainer, not re-quoted.
Exit	Your historical data exported in full on request. No lock-in, no export fee.

Why teams pick Actowiz for this work

Engineers, not a dashboard. You get people who fix breakages, not a self-serve tool you maintain yourself.
We tell you what we can't do. Scope limits and coverage gaps are stated before you sign, not discovered in month three.
QA is part of the service. Schema validation and sampled human review run before delivery, every run.
Compliance is documented. Sources, method and lawful basis written down so your legal team can review them.
Fixed monthly cost. No per-request metering, no surprise overage on a month when a competitor adds SKUs.
Six years, 40+ countries. Long-running production pipelines across retail, travel, mobility and finance.

Definitions

Terms used on this page

Plain definitions of the terms used on this page, so procurement and legal reviewers are working from the same vocabulary as your data team.

FAQ

Agent feeds: frequently asked questions

Straight answers for teams building on this.

Same collection, different shape. The catalogue feed is built for analysts and dashboards; this is built for a machine that cannot fill gaps by judgement.

The differences that matter: attributes are typed and unit-normalised, absence is explicit with a reason, and freshness is stamped so the agent can judge staleness. Most catalogue feeds fail all three.

Because an agent that cannot tell 'the maker says no' from 'nobody said' will exclude products that might qualify and tell the user they do not — asserting something nobody claimed.

That is the failure mode that destroys trust in an agent fastest, and it is entirely a data problem rather than a model problem.

No, deliberately. Embeddings should be generated against your model on your schedule, and a vendor's vectors go stale the moment you change models.

We deliver structured fields. The vector layer, the ranking and the prose generation are your product.

We can, and we will not put the result in an observation field. If you want inferred attributes, run the model on our data and keep the output in a column that is visibly a prediction.

The moment we generate an attribute, your provenance chain breaks and you can no longer tell a user or a merchant which facts came from the manufacturer.

It depends on your promise, which is why we set safe_to_recommend_until per field group rather than one number. Price and availability usually need a much shorter horizon than attributes.

The point is that the agent can see how old a record is and degrade to 'check availability' rather than recommending confidently from stale data.

You will see it in the sample, per attribute, before committing. Fashion and toys are typically thinner than electronics and industrial.

Thin coverage is not automatically fatal — an agent that asks a clarifying question is better than one that guesses — but it changes what you build, and it should change it before launch rather than after.

Yes, and for a discovery agent it usually must. Cross-retailer records arrive with a match confidence so weak links can be excluded by your own threshold rather than ours.

For a brand-side agent the same data doubles as a discoverability audit — showing where your catalogue cannot answer constraints that competitors can.

Yes, and it is where most of the demand we see is coming from. The additional considerations are pincode-level availability, which our quick-commerce work already handles, and multilingual titles, where we retain the original and do not machine-translate into the observation field.

We quote individually. Drivers are catalogue size, attribute breadth, source count and refresh frequency — refresh matters most because a short safe-to-recommend horizon means more observations.

One scoping call, a free sample within 24 hours with per-attribute fill rates, then a fixed monthly quote. Request a quote.

Run your agent against a real sample first

Tell us the constraints it must answer. We return real records with typed attributes, reasoned nulls and per-attribute fill rates.

Get a free agent-feed sample Book a 20-min scoping call

No sales sequence. If your key constraint attribute is thin in your category, the sample is where you find that out rather than after launch.

Enterprise Data Extraction

AI Training Data

24-hour sample

AI Dynamic Pricing

Social Commerce

GCC Quick Commerce

KitchenIntel

UK Grocery Price Tracker

What we can collect

Sample data

Ready-Made Scrapers

API Playground

Need it managed instead?

Case Studies

Sample Datasets

ROI Calculator