Product Feed Optimization: Your Feed Is Now What AI Reads First

Product feed optimization used to be a Google Shopping chore owned by whoever ran paid. The same feed is now the most structured description of your catalog that AI assistants and shopping agents read when deciding what to recommend and, increasingly, what to transact against. This covers the attributes that matter, a feed audit you can run today, the failure modes that break feeds silently, and the check almost nobody runs.

Who this is for: heads of ecommerce, growth and catalog operations at brands with more than a few hundred SKUs.

What changed

For a decade the feed was a means to an end: get products into Google Shopping, keep disapprovals down, let bidding do the work. Two things broke that model.

Discovery fragmented. A meaningful share of product research now begins in ChatGPT, Gemini, Perplexity or a Google AI Overview. Those surfaces do not crawl your site the way a search engine did. They assemble an answer from structured sources, and your product feed is the most structured source you publish.

The shopper became a shopper plus an agent. Agent traffic, meaning AI shopping on a buyer's behalf, behaves nothing like human traffic. It does not scroll, does not load your hero image, and does not wait for a carousel. It reads structured data and leaves. If the answer is not in the structured data, the answer is that you were not recommended.

This is why feed work now sits with AI visibility rather than under paid media, and why it is the substrate of the whole discovery problem. Nudge Catalog Enrichment exists for exactly this: making every product understandable, recommendable and transactable by AI, then pushing those changes to the surfaces that agents actually read. Feed optimization is the visible half of that job.

A useful way to hold it: being recommended is the start, being bought is the win. Your feed decides whether you are in the recommendation at all.

What a product feed contains

A structured file, usually XML, TSV or a direct API sync, with one row per sellable variant. Think of the attributes in three tiers.

Tier 1: required, or the item does not appear

Tier 2: competitive, or you appear for fewer queries

brand, gtin, mpn, condition, product_type, google_product_category, item_group_id, color, size, material, age_group, gender, shipping, sale_price.

Two deserve attention. gtin is how surfaces confidently match your item to a known product, which unlocks comparison placements, and missing GTINs on branded goods is a quiet, expensive gap. item_group_id tells a surface that eleven rows are eleven colorways of one product rather than eleven unrelated items. Get it wrong and you compete against yourself.

Tier 3: agent critical, and mostly ignored

This is the tier that changed. When an assistant answers "which of these is best for X", it needs the attributes that let it reason about fit, not just identity. Material composition. Care instructions. Dimensions with units. What the product is actually for. Return window. Warranty terms.

Most feeds carry the identity attributes and skip the reasoning attributes, because identity was all Google Shopping strictly needed. Catalog Analysis in Nudge scores each SKU on exactly this gap, so you get a per product view of what AI can and cannot understand rather than a feed level pass or fail. That distinction matters, because a feed can be 100 percent valid and still be unreadable in the way that decides a recommendation.

Google Shopping optimization: the parts that pay

Title structure. The highest leverage field in the file, and it is not your page headline. Front load, because titles get truncated. A durable pattern:

Brand + Product Type + Key Attribute + Secondary Attribute

So Ursa Major Fantastic Face Wash 120ml Gentle Daily Cleanser beats Fantastic Face Wash. Stay within 150 characters, put the terms a buyer would type in the first 70, and never pad with "Free Shipping" or "Best Seller". Those get items disapproved and teach the matching engine nothing.

Category taxonomy. google_product_category is a controlled vocabulary. Guessing, or letting a platform auto assign, is one of the most common reasons a technically valid feed underperforms. Assign the most specific applicable node, not the safest broad one.

Image discipline. Main image on a clean background, no logos, no text overlays. Lifestyle belongs in additional_image_link.

Price and availability freshness. If the feed updates nightly but inventory moves hourly, you will show out of stock items and get suppressed. Move to a fetch measured in hours, or an API sync, on any catalog with real velocity.

The check almost nobody runs: does your feed agree with your page?

Here is the failure mode no feed tool flags, because feed tools only look at the feed.

Your feed says in stock at $48. Your product page renders $52 after a price rule, and the size the buyer wants is greyed out. A human notices and adjusts. An assistant reading both sources sees a contradiction and does the safe thing, which is to recommend something else. An agent trying to transact hits a mismatch and abandons.

Feed and page divergence is invisible to every tool that audits only one of them, and it happens constantly, because the feed and the page are usually generated by different systems on different schedules.

The fix is to treat the feed, the on page structured data and the rendered page as one artifact that must state the same facts. This is the parity guarantee in Nudge Catalog Enrichment: the version of a page built for agents is provably fact identical to the one a shopper sees, just machine legible. Same price, same stock, same variants, same claims, different formatting. Optimized Storefronts maintains that agreement continuously rather than leaving it to a quarterly cleanup, and every change is one approval, landing on your own domain and in your own Merchant Center account so authority accrues to your brand rather than a vendor's hosted copy.

You can approximate the check manually. Pick twenty SKUs, pull the feed row, pull the JSON-LD from the page, load the page, compare price, availability and variant list across all three. Zero mismatches on twenty means your pipeline is healthier than most. Most teams find several.

Where feed work sits: coverage

Verify competitor rows against current vendor documentation before publishing.

A feed audit you can run this afternoon

1. Coverage. Percentage of sellable SKUs present in the feed. Compare feed rows to active variants. Under 95 percent needs an explanation.

2. Disapprovals and warnings. Sort by item count, not severity. One systemic warning across 4,000 items beats four critical errors on four items.

3. GTIN coverage. Percentage of rows with a valid GTIN, broken down by brand or category.

4. Title audit. Sample 50. How many lead with brand? How many exceed 150 characters? How many duplicate another row?

5. Category assignment. How many rows sit on a top level node rather than a leaf? Those are underspecified.

6. Freshness. Time between an inventory change in your commerce platform and that change appearing in the feed. Measure it, do not assume it.

7. Parity. The twenty SKU comparison above.

Write the seven numbers down. That is your baseline, and it is what you re-measure in ninety days. Catalog Analysis produces checks 1 through 6 continuously and check 7 is the parity guarantee, but the manual version is genuinely useful and costs you an afternoon.

What breaks feeds silently

• Re-platforming changes id values. Performance history resets and nothing tells you.

• A category tree refactor shifts product_type and quietly changes what you match against.

• Variant attributes inherited from the parent, so every colorway reports the same color.

• Truncation of a description that is fine at 400 characters and gets cut mid sentence by a surface limit.

• Currency and locale drift on multi region feeds, where the price is right and the currency code is not.

• A sale price that expires but sale_price_effective_date does not.

None of these throw an error. They just make you less recommendable.

From feed to revenue

A clean feed gets you into the recommendation. It does not finish the sale. That is the second moment, and it is where most feed programs stop measuring.

The loop worth running is Insights, then Opportunities, then Revenue. Nudge AI Visibility shows which shopping questions you appear in and which you do not, down to individual SKUs. Catalog Enrichment fixes what the feed and catalog are missing for the products that should be appearing. Nudge Funnels then generates the page each of those clicks deserves, on your own domain, so a visitor arriving from an assistant answer meets a page built for that context rather than a product page built for everyone.