FeedManager
Language: English

AI

AI that fixes product data — and shows its work

Fill the gaps channels reject products for, write titles and descriptions that sell, and let AI set up the tedious parts of a feed. Every suggestion is reviewed before it is used, and your source data is never overwritten.

Request an invitation What it costs

Suggestions, not surprises

Your value next to the AI's, until you decide

A run shows the source value next to the suggestion with a confidence score. Accept per product, per page or everything above a confidence — and undo a field or a whole run later.

  • Aim a run with the product list's own filters, for example "only where the colour is empty", or at tags such as Best sellers.
  • See the exact number of credits before it starts.
  • Approved values go to the feeds that map them; the affected feeds rebuild automatically.
Attribute completion 3 of 1,284 products
color · Trail Runner GTX
—
Slate grey
94%
material · Linen Shirt Relaxed
—
Linen
97%
gender · Kids Rain Jacket
—
unisex
71%
An example of the review screen. Nothing reaches a feed until it is approved.

What it does

The jobs a feed manager would rather not do by hand

Complete the catalog

Google product category

Picks a category from Google's official taxonomy — never an invented one — so shopping campaigns and Snapchat get the category they need.

Missing attributes

Reads colour, size, gender, age group, material, pattern and condition from the product's own title and description, in one pass.

Attributes from the image

Colour, material and pattern taken from the main image, only where neither the source nor an earlier suggestion has them.

Write the content

Titles

Rewritten for Google Shopping, Meta, marketplaces or price comparison sites — each with its own length limit and what it puts first. With Search Console connected, each product's five most clicked search queries join the prompt as wording hints.

Descriptions

Plain-text descriptions of about 300, 800 or 1,500 characters, with paragraphs and no HTML, from the facts the product already has.

Translations

Titles and descriptions in English, Turkish, German, Spanish, French and Italian, with your own glossary for the terms that must not change.

Your brand voice, and hard limits on every text

Set the tone, the capitalisation, words never to use and a disclaimer we append word for word. Whatever the model writes is then checked: the channel's length limit, no web or email addresses, none of your banned words, and no promotional claims — "sale", "free shipping", "cheapest" — that your own text didn't make. A text that fails is dropped and not charged. The same word lists drive the wording checks of your feeds, so the texts AI writes and the texts feeds send follow one set of rules.

Check the images

Image check

We measure every image ourselves, free: unreachable, not an image, smaller than 100 pixels (a disapproval) or under 500 pixels on the short side. The model then looks for promotional overlays, placeholders, busy backgrounds on the main image, a picture that doesn't match the product, or several products in one picture.

The best main image

When a product has several images, the model can suggest a better one as the main image — always one of the product's own. Filter the product list or a feed on the findings, for example to leave out products with image problems.

Ad images with prices and discounts on them are made by Creative Studio (Enterprise) — drawn from templates and your product data, without AI.

Set up the feed

Mapping suggestions

Proposes which of your fields fills which field of the catalog or the channel, from field names and example values. Empty rows are filled; rows you mapped differently get a suggestion to accept or ignore.

Rules in plain language

"Add the brand in front of titles that don't start with it" becomes rules in the editor, marked as AI, checked against your fields — and saved only when you save.

Custom-label schemes

Price bands, best sellers from GA4, stock depth, margin or category groups as custom_label rules — worked out from statistics of your catalog, not from individual products.

Value normalization

"Navy", "navy blue" and "NAVY" become one value: up to 1,000 spellings of a brand, colour, size, material or pattern turned into a lookup table you review first.

Feed errors explained

When a build leaves items out — or Google reports issues in Merchant Center — the model explains them in plain language and proposes a fix. It changes nothing by itself.

Catalog health

A free report of which of 16 fields your products are missing — ranked by the products that get views when GA4 is connected — with a link to the action that fixes each gap, and an optional AI plan.

How feeds use it

AI values sit next to your data, not on top of it

An approved value is stored beside the product, re-applied after every sync and used only where a feed asks for it.

  1. 01

    Choose and quote

    Pick an action, the products and the model. The quote shows the credits before anything runs; small runs answer in seconds, big ones run in the background.

    • Up to 100,000 products per run
    • Only products still missing the field
    • Answers we already have are free
  2. 02

    Review

    Each suggestion shows the source value, the AI value and a confidence. Keep what is right, discard the rest, or approve everything above a threshold.

    • Nothing offered below 40% confidence
    • Compare two runs product by product
    • The share of suggestions you kept
  3. 03

    Map and publish

    Map {ai.title} or {ai.color} in a feed and it uses the AI value where there is one and the source's value everywhere else.

    • Opt in per feed and per field
    • Undo a field or a whole run
    • Channel validation still runs after AI

Automation and control

Hands-off when you want it, never out of your hands

Autopilot

Once a day, run the actions you choose on products that are new or changed since the last run — optionally only those with tags you pick — within a monthly credit limit you set. It never runs over the whole catalog.

Auto-approve

Off by default. Per action, apply suggestions at or above a confidence you choose; for generated text you also confirm that it may reach feeds unread.

Choose the model

Economy, Balanced or Premium, or one exact model — for all actions or per action, such as Premium titles and Economy categories. Block a provider you don't want.

Compare two models

Run the same action on up to 20 products with two models, compare the answers side by side and make the better one the default for that action.

Your own API key

On Premium and Enterprise, bring your own key from Anthropic, OpenAI or Google: runs then use only the providers you have keys for and cost 20% of the usual credits.

Undo

Remove what a run wrote for one field or all of them; values a later run replaced are left alone, and the products go back to the source's values.

Or ask your own AI assistant

Claude, Cursor and other agents can connect to FeedManager through the Model Context Protocol: read your catalog, feeds, Merchant Center issues and readiness, start a sync, a build or a page check and tag products — with an API key you control, on Pro and above. Their own reading and reasoning run on your assistant's subscription, not on FeedManager's AI credits. Only when you give the key the ai:write scope can an agent ask FeedManager's AI about a product; that spends your credits, up to a daily limit you set for the key. Set up MCP

Safety

AI can't touch what it shouldn't

Product feeds are untrusted text, and a price is not a matter of opinion. The limits are built into the system, not left to the prompt.

  • AI can only fill descriptive fields. Prices, identifiers (GTIN, MPN, SKU), product links and availability are not on its list, so no text in a feed can make it change them — and a suggested main image can only be one of the product's own images.
  • Answers must match a fixed shape: a category outside Google's taxonomy, or a gender, age group or condition outside the allowed values, is dropped.
  • We download product images ourselves, with the same private-address checks as every fetch, and send them to the model directly — no AI provider fetches your URLs.
  • Search queries from Search Console are words strangers typed. Title runs pass them to the model as hints only — nothing the product data doesn't support, no shop or competitor brand names — and they are never available to rules or templates.
  • Models from three providers: Anthropic (Claude), OpenAI and Google (Gemini). The panel names the company next to each model, and you can block any of them.
  • FeedManager does not train AI models on your data.
  • Every run, approval, purchase and settings change is in the activity log.

Credits

Pay for answers, not for tries

AI runs on prepaid credits. The quote is exact, a cached answer is free, and anything that failed, was skipped or produced nothing usable goes back to your balance.

AI credits included per plan
PlanCredits includedBest model tier
Free 200 once Balanced
Standard 2,000 a month Balanced
Pro 10,000 a month Premium
Premium 20,000 a month Premium
Enterprise Custom Premium
  • The monthly credits of a plan expire at the end of the month; they are used first.
  • Need more? Credit packs from $25 for 10,000 credits, valid 365 days — with reminders before any credits expire.
  • A category costs 2 credits, a title 3, a description 5 at the base rate; Economy models cost less, Premium models more.
Credits per action and packs

AI

Questions about AI

Does AI change my store or my source data?

No. A suggestion is stored next to the product and only becomes part of a feed when you approve it and the feed maps it, for example as {ai.title}. The source's value stays, is still used everywhere else, and comes back if you undo the suggestion. FeedManager never writes to your store.

Which AI models does FeedManager use?

Models from Anthropic (Claude), OpenAI and Google (Gemini) in three tiers — Economy, Balanced and Premium. By default FeedManager picks a suitable model per action; you can choose a tier or one exact model for all actions or per action, block a provider, and compare two models on the same products. Free and Standard can use up to Balanced models, Pro and above up to Premium.

Can AI output reach my feeds without a review?

Only if you allow it. Auto-approve is off by default; you can switch it on per action above a confidence between 40 and 100, and for generated text you must also confirm that it may reach feeds unread. Channel validation still checks every item after AI.

Can AI change my prices or product identifiers?

No. AI can only fill descriptive fields such as the category, colour, material, title and description. Prices, GTIN, MPN, SKU, product links and availability are outside what it may write, so no text in a feed can make it change them.

Is my data used to train AI models?

FeedManager does not train AI models on your data. For a run, the product text — and for image actions the images, which we download and send ourselves — goes to the provider of the model that answers it, which you choose and can restrict.

Which languages can AI write in?

Texts are written in the product's own language by default, or one you choose. Translations cover English, Turkish, German, Spanish, French and Italian, with a glossary per language for terms that must not change.

Can I use my own API key?

Yes, on Premium and Enterprise. With a key from Anthropic, OpenAI or Google, your runs use only the providers you have keys for, the provider bills you for the model, and a run costs 20% of the usual credits.

Let AI fill the gaps in your catalog

Every plan includes AI: the Free plan starts with 200 credits, paid plans get a monthly allowance.

Request an invitation Compare plans

Private beta — accounts by invitation.