How do I get products recommended in AI answers and AI Overviews

You get products recommended in AI answers by making the SKU a fact an engine can defend: unique specs in crawlable HTML, Product structured data that matches the visible PDP, reviews and brand identity that do not contradict each other, and a Merchant Center feed that agrees with the page. Google can include product listings inside generative responses; ChatGPT shopping is a partner feed; Perplexity shopping is a merchant index plus ordinary web citations. None of those surfaces publish a “pay this field, win the card” formula.

This spoke sits under the Answer Engine Optimization playbook. Ranking the PDP and missing the Overview is a different diagnostic — use ranked but missing AI Overviews. Being named without a product card is a mention vs citation problem, not a feed problem.

The short answer

• Treat “recommended” as three jobs: Google product listing inside an AI response, ChatGPT shopping/discovery, and Perplexity recommendation or citation. Do not run one playbook as if they were one API.
• Google’s own AI-features and AI-optimization pages tell you to keep Merchant Center and Business Profile current, match structured data to visible text, and write non-commodity content. They also say structured data is not required for generative AI search.
• Unique specs on the PDP beat synonym titles. A model comparing “best 16-inch travel backpack under 2 lb” needs weight, dimensions, and materials in text, not in a lifestyle hero.
• Reviews and entity identity (brand, GTIN, same product name everywhere) decide whether the SKU is one thing or a cloud of near-duplicates.
• Measure a frozen SKU panel weekly. I will not invent a conversion lift for your catalog. SEO certified since 2021 does not make Overviews kinder; it makes the data layer boring to audit.

What does “recommended” mean on Google vs ChatGPT vs Perplexity?

“Recommended” is a shopper word. The engines are doing different jobs with different inputs. If you collapse them into one KPI, you will “fix schema” for a ChatGPT miss that was never reading your JSON-LD.

Surface	What “recommended” looks like	Primary input Google / the vendor documents	What they have not published
Google AI Overviews / AI Mode	Product facts or listings inside a generated Search response, plus supporting links	Indexed, snippet-eligible pages; Merchant Center and Business Profile kept current; structured data matching the page	A Product schema type that forces Overview inclusion
Google Shopping tab / free listings	Classic shopping units	Merchant Center participation is required for the Shopping tab	That the Shopping tab equals an Overview citation
ChatGPT shopping / discovery	Product cards, comparisons, or cited shopping answers	Structured product feed for approved ACP partners	That on-page Product JSON-LD is the ChatGPT catalog
Perplexity answer	Named SKU with a citation to a URL	Live web retrieval plus citations	A public ranking formula for which PDP wins
Perplexity Shopping	Product cards / “recommended product” in the shopping experience	Merchant Program catalog share (their blog: more complete details in their index)	Weights, review thresholds, or a paid placement product for those cards

Operator rule: pick the surface before you pick the ticket. A Merchant Center outage will not explain a Perplexity citation that still links a Wirecutter roundup.

Checklist before you brief merchandising:

• Name the surface: Overview, AI Mode, ChatGPT shopping, Perplexity answer, Perplexity Shopping
• Name the SKU and the shopper prompt (not a category keyword)
• Screenshot the unit: listing, citation, mention-only, or absent
• Record whether the buy link is your PDP, a marketplace, or a third-party review
• Do not start a schema sprint until that row exists

What does Google actually document for products in AI features?

Google’s AI features page is explicit: eligibility for a supporting link is ordinary Search eligibility — indexed, snippet-eligible, no extra technical bar. The same page’s best-practice list includes making important content textual, matching structured data to visible text, and checking that Merchant Center and Business Profile information is up to date.

The generative AI optimization guide is the product-specific sentence merchants skip: generative AI responses can include product listings, product information, and local-business information, and using Merchant Center (including feeds) and Google Business Profiles can help products and services show in both AI responses and other Search results. That is a “can help,” not a ranking guarantee.

Same guide, mythbusting section you should tape to the sprint board:

Claim you will hear	Google’s documented position (as of August 2026)
“Ship a special AEO schema type”	There is no special schema.org markup required for generative AI search
“JSON-LD is required to appear in Overviews”	Structured data is not required for generative AI search; it remains useful for rich results
“Write llms.txt so Overviews can shop the catalog”	Google Search does not use special AI text files as a ranking or inclusion switch
“Buy inauthentic mentions so products get named”	Inauthentic mentions are not a helpful strategy; spam systems still apply
“Chunk PDPs into FAQ chips for AI”	No requirement to fragment pages for AI understanding

So the honest Google job is: stay eligible, keep the commerce graph honest, and write specs a comparison query can lift. UCP and in-chat checkout are a later surface. Google’s optimization guide mentions Universal Commerce Protocol as an emerging way for Search agents to do more, and the retailer announcement describes checkout inside AI Mode and Gemini for eligible merchants. That is not “Product schema got us the Overview.”

Procedure for the Google path:

1. Confirm the PDP is indexed and snippet-eligible (URL Inspection, no nosnippet / max-snippet:0 on the spec block).
2. Confirm Merchant Center has the SKU, with title, price, availability, and identifiers that match the page.
3. Confirm Business Profile facts if you also sell local or in-store.
4. Confirm JSON-LD Product matches the HTML the shopper sees.
5. Only then rewrite copy so the unique specs are in the first screen of text.
6. Log the prompt in Search Console’s generative AI performance report if the property has it.

If the page cannot appear as a blue link, it is not a product-recommendation problem yet.

What does Product schema buy you — and what does it not?

Product on schema.org describes a named SKU or packaged offering. Google splits that type into two Search experiences: product snippets (reviews, roundups, pages that are about a product you may not sell) and merchant listings (pages where a shopper can buy from you). Retail PDPs should build to merchant listing, not the lighter snippet floor.

What it buys, per Google’s own ecommerce data page: better eligibility for product rich results, more accurate extraction of price / discount / shipping, and a cleaner check when Merchant Center verifies the feed against the site. What it does not buy: a dedicated Overview ranking signal. Google says structured data is not required for generative AI search and warns against over-focusing on it.

Job	Product JSON-LD helps	Product JSON-LD does not do
Rich results (price, availability, stars)	Eligibility when required properties validate	Guarantee the unit shows
Merchant Center verification	Site and feed can be compared on the same fields	Forgive a lying feed
Shopping tab	Can assist verification	Replace Merchant Center participation
AI Overviews / AI Mode	Keeps facts machine-consistent with the page	Force a product listing inside the Overview
ChatGPT catalog	Indirect, only if a crawler later reads the PDP	Substitute for the ACP product feed
Perplexity Shopping index	Possible if they retrieve the PDP	Replace their Merchant Program feed

Google’s merchant-listing technical rules that teams skip:

• Only purchasable pages on your site. Affiliate hop pages are out.
• One product (or variants of the same product) per URL. “Shoes in our shop” is not a Product.
• Distinct URL per currency if you sell in more than one.
• Put Product in the initial HTML if you care about shopping crawls. JavaScript-generated Product markup can make Shopping crawls less frequent and less reliable — Google’s words, especially for price and availability.

Checklist:

• @type is Product (add a co-type like Book only when it is true)
• Required merchant-listing fields present: name, image, nested Offer with price and priceCurrency
• Offer is Offer, not AggregateOffer, on a merchant PDP
• Markup is in first HTML, not only a client render
• Rich Results Test is clean for the live URL, not a staging clone
• You are not marking up a category template as one SKU

Search Console splits the aftermath into two reports, because the requirements differ: the Merchant listings report for pages where a shopper can buy, and the Product snippets report for review / aggregator pages. Do not debug a PDP in the snippets report and call the listing “fine.”

Report	Use it for	Ignore it for
Merchant listings	Buyable PDPs with Offer	Editorial roundups you do not sell
Product snippets	Review pages, “best of” posts	Your cart URLs
Rich Results Test	Live URL validation	A plugin preview on localhost
URL Inspection	Fetched HTML vs your CMS	“It looks right in Chrome”

Variants: Google wants markup on product pages, and documents a separate product variant pattern (ProductGroup with variesBy, hasVariant, productGroupID, plus inProductGroupWithID / isVariantOf on the child). That markup is for merchant listing variant display. It is not a documented Overview inclusion switch.

Site pattern (Google’s two designs)	Markup job	Recommendation risk
Single page, variants via query params, no reload	Markup all buyable variants; keep the fetched HTML honest for the default offer	JS swaps price while JSON-LD stays on the parent
Multi page, each variant has a URL	Mark each URL as its own Product in the group	Category URL marked as one SKU
One URL, one offer, color as an image	You are selling one object	Comparison prompts cannot pick a size

A valid JSON-LD blob on a hollow PDP is a dressed-up empty box. The Overview still has nothing unique to recommend.

Which Product fields actually matter on a PDP?

Build to Google’s merchant-listing required set, then fill the identifiers and attributes a comparison query needs. Required for merchant listings: name, image, and a nested offers Offer. On the offer, merchant listings require a price greater than zero and a currency. Availability, GTIN, brand, MPN, SKU, color, size, material, shipping, and returns are recommended — which in practice means “skip them and you are technically valid and commercially mute.”

Field	Why a recommendation engine needs it	Failure if it is missing or wrong
name	The string the answer will copy	Title spam (“best 2026 deal”) becomes the quoted name
image	Shopping cards and image-grounded Search	Broken or lifestyle-only images fail crawlability
offers.price + priceCurrency	Price in the unit	Stale sale price vs PDP
availability	Whether the SKU is a live option	Recommended then 404s at cart
brand.name	Entity join	Your house brand vs manufacturer brand fight
gtin* / mpn / sku	Join feed, page, and marketplace listings	Duplicate “products” for one object
color / size / material	Attribute filters in comparison prompts	Generic “backpack” with no discriminators
aggregateRating / review	Social proof in cards when policies allow	Fake stars = policy risk, not a ranking cheat
shippingDetails / hasMerchantReturnPolicy	Offer completeness	Eligible for listing enhancements; not an Overview switch

Procedure for a 50-SKU sample (do not boil the ocean on day one):

1. Export the live PDP HTML and the JSON-LD.
2. Diff name, price, currency, availability, brand, and GTIN against Merchant Center.
3. List attributes that appear in shopper prompts (weight, battery life, ISO, inseam) and check they exist as visible text, not only as additionalProperty hope.
4. Fix mismatches before you write new marketing copy.
5. Re-test with URL Inspection and the Merchant Center diagnostics, not a local schema plugin screenshot.

If the unique spec only lives in a spec-sheet PDF, you taught Google a brochure and hoped the Overview would open the attachment. It will not.

additionalProperty on schema.org is a dump drawer. It can carry a custom name/value, and some teams hide the real differentiators there while the HTML stays “premium, durable, versatile.” Google’s AI-features page still wants important content in textual form. Treat additionalProperty as a mirror of a visible row, not as the only copy of the fact.

Procedure for one comparison prompt:

1. Write the shopper sentence out loud (“best 16-inch backpack under 2 lb with a laptop sleeve”).
2. List the attributes that sentence requires (size, weight, sleeve dimension).
3. Find each attribute on the PDP as visible text with units.
4. Copy the same values into Merchant Center attributes and into JSON-LD fields Google actually documents (size, material, color) when they fit.
5. If a required attribute has no Google field, keep it in the HTML table anyway. Do not wait for a new schema property to tell the truth.

Why unique specs beat a synonym catalog?

Google’s AI-optimization guide says unique, non-commodity content is the long-run job, and that first-hand experience beats a summary of what already exists. A PDP that restates the manufacturer’s three adjectives is commodity. A PDP that states measured weight, packed volume, warranty terms, incompatible use cases, and what you actually changed in this revision is a fact a comparison answer can lift.

Query fan-out makes this worse. Google documents fan-out as extra related searches used to build the response. “Best carry-on for a 14-inch laptop” can spawn size, weight, and laptop-sleeve sub-queries. If your page never states interior dimensions, the Overview will cite the retailer who did.

PDP pattern	What the shopper / model gets	What to ship instead
Title + three lifestyle bullets	Adjectives	Numeric specs in a table
Variant soup on one URL	Ambiguous offer	One URL per buyable variant, or documented variant markup
Specs in images only	Alt-text hope	Spec table in HTML
“As seen in” with no numbers	Prestige fog	Dated measurements you can defend
Category copy pasted onto 40 SKUs	Duplicate cluster	Unique constraints, accessories, and not-for list per SKU
Specs behind a tab that never renders	Empty fetch	Specs in the initial HTML

Unique does not mean a 2,000-word brand essay. It means the attributes the prompt is comparing exist as text on this URL.

Checklist per money SKU:

• Spec table with units (in / cm, lb / kg, Wh, ISO)
• Who it is for, and who should skip it, in two short lists
• What’s different from last year’s SKU or the obvious competitor, in facts not slogans
• Compatibility (what it fits, which chargers, which bags) in text
• Date on any number that can go stale (firmware, included accessories)
• Same numbers in JSON-LD / feed attributes where Google has a field (color, size, material, shipping)

I will cite receipts I can defend on a sales call: SEO certified since 2021, hundreds of production sites shipped. I will not invent “SKU A gained 18% Overview share after we added a spec table.” If you cannot say the number out loud, do not put it in the passage you hope gets lifted.

How should reviews and entity identity work together?

A product is an entity. Reviews are evidence about that entity. If the brand string, GTIN, and product name disagree across the PDP, the marketplace listing, and a review widget, you handed answer engines three objects.

Google’s merchant-listing review rules: follow the review snippet guidelines. Reviewer name must be a real Person or Team, not “50% off on Black Friday.” Markup must match visible reviews. Self-serving stars you invented in JSON-LD are a policy problem, not an AEO tactic.

Off-site reviews still matter for engines that retrieve the open web. Perplexity’s shopping posts describe synthesizing pros/cons from customer reviews. That is their product copy, not a published weight. Google’s AI-optimization guide also warns that chasing inauthentic mentions is not a helpful strategy. Earn real reviews and real editorial tests. Do not buy a comment farm and call it entity work.

Signal	Healthy	Broken
Product name	Same string on H1, schema name, feed title	Marketing title vs feed title vs Amazon title
Brand	One Brand name that matches packaging	Manufacturer in schema, house brand on the H1
GTIN / MPN / SKU	Join page, feed, and marketplace	Missing GTIN so the same object fragments
Aggregate rating	Visible widget = marked-up numbers	Stars in JSON-LD only
Review authors	Named people or a named team	Coupon text as author
Third-party tests	Independent measurements that match your table	Affiliate roundups you secretly wrote

Procedure:

1. Freeze the canonical product name (the one on the box).
2. Map GTIN + MPN + internal SKU in one sheet.
3. Make H1, schema name, and feed title describe the same object. Marketing flavor can live in description, not in the identity fields.
4. Only mark up reviews that render for Googlebot.
5. If you sell on marketplaces, treat those titles as a fourth copy of the entity — they will get retrieved too.

Failure mode: the Overview recommends “Acme Pack 20” and the click lands on a 2024 colorway with a different GTIN. The shopper did not get a recommendation. They got a collision.

How do Merchant Center feeds actually help AI answers?

Google’s share your product data page is the clean split: structured data on the page, feed in Merchant Center. Uploading a feed is not mandatory to appear in Google Search results. It is mandatory for some surfaces, including the Google Shopping tab. Google Search may use Merchant Center data for product rich results. The AI-optimization guide is the extra clause: feeds can help products appear in AI responses.

That is the hedge. Merchant Center is the commerce graph Google already trusts for shopping. Generative features can draw on it. It is not a documented “feed completeness score → Overview slot” function.

Feeds also buy operational control Google crawling cannot: confidence that Google knows the full catalog, update timing you choose (including Content API for stock), and data that is not on the website (store-level inventory).

Experience (Google’s table)	Structured data	Merchant Center
Product rich results	Used	May be used
Google Images product annotations	Used	Images listed in Merchant Center
Google Shopping tab	Can help verification	Required
Generative AI product listings	Not a separate markup type	Documented as a way to help products show in AI responses

Feed hygiene that actually shows up in AI answers:

• Every money SKU has id, title, description, link, image_link, price, availability
• Identifiers: gtin and/or mpn + brand per Merchant Center product data spec
• link is the canonical PDP, not a tracking wrapper that 302s into a maze
• Automatic item updates enabled so the site can correct a stale feed price or stock (Google’s sync warning)
• Free listings enabled if you expect organic shopping units at all
• Disapproved items triaged weekly — a disapproved SKU is not “waiting on AEO”

How Google says to keep the feed honest when the site moves faster than the file:

Update method	Use when	Hedge
Automated feed from crawled pages	Small catalog, slow price changes	Crawling is not guaranteed to find every product
Scheduled feed file	You need control over when Google gets the snapshot	Still lags between uploads
Content API	Stock and price that change intra-day	Still must not fight the PDP
Automatic item updates	Feed and HTML diverge on price or availability	Google’s documented sync fix, not an AEO plugin

Procedure:

1. Pick a SKU that sold out yesterday.
2. Note PDP availability, schema availability, and Merchant Center availability at the same minute.
3. If any of the three still says in stock, you do not have a recommendation program. You have a lie with an SLA.
4. Turn on automatic item updates unless you have a reason the site is the wrong source of truth.
5. Re-fetch after the next feed / crawl cycle and log the delay. Do not promise Overviews “same day.”

Do not wait on UCP to do this. UCP, in Google’s own UCP guide, uses existing Merchant Center shopping feeds and requires an active Merchant Center account plus checkout-eligible products for their implementation. If the feed is dirty, checkout protocols will inherit the dirt.

How does ChatGPT shopping differ from an Overview citation?

ChatGPT shopping is not Google Search. OpenAI’s Agentic Commerce get-started page says product feeds give ChatGPT the catalog data it needs to index products, understand attributes, and present accurate information in shopping experiences. Onboarding those feeds is currently available to approved partners. File upload (typically a daily snapshot) plus API updates is the documented path. Promotions data is API-only.

That is the hedge: if you are not an approved partner, shipping prettier Product JSON-LD on the PDP does not enroll the catalog. The PDP can still be retrieved if ChatGPT Search cites the open web — that is a citation problem, not a shopping-feed problem. Track them as different rows.

OpenAI’s feed best practices are unglamorous on purpose: factual descriptions, valid encoded URLs, stable product id plus unique variant id, variant-specific title/url/price/availability when those differ. Optional HTML/markdown descriptions can improve answer quality and are not required for ingestion.

ChatGPT path	What it is	What it is not
ACP product feed (approved)	Catalog Google-style: title, images, price, availability, variants	A guarantee you will be the recommended SKU
ChatGPT Search citation of your PDP	Ordinary web retrieval	Proof the shopping graph knows you
Marketplace listing ChatGPT already knows	A third party’s feed	Your brand’s facts winning
In-chat checkout	Partner features that have changed since 2025	Something to promise your CFO from a blog recap

OpenAI also publishes a prohibited-products policy on that get-started page (adult, age-restricted, weapons, prescription-only, and similar). If the category is blocked, feed quality will not save it.

Procedure if shopping recs are the actual goal:

1. Confirm whether you are in the approved-partner program. If not, stop selling “ChatGPT schema work” as shopping enrollment.
2. If yes, map SKUs to ACP fields; do not invent columns.
3. Snapshot daily; upsert price and availability through the day if they move.
4. Keep seller.name as the name the shopper should see.
5. Log ChatGPT prompts separately from Google Overview prompts. Same SKU, different systems.

I am not going to recap third-party Instant Checkout obituaries as if they were your KPI. Checkout features have moved. Discovery via feed is what the current developer docs describe. Hedge checkout; ship the catalog.

How does Perplexity shopping differ from a cited web answer?

Perplexity can recommend a product in two different ways that teams mash together. One: an ordinary answer with citations to PDPs, reviews, and forums. Two: the Shopping experience with product cards. Perplexity’s own Shop like a Pro post launched a Merchant Program so large retailers can share product specs and live details. They say the program is free, distinct from sponsored-question ads, and that joining increases chances of being a “recommended product” because the products will be in their index and more complete details help them judge quality and relevance. That is their marketing sentence. They did not publish a ranking algorithm, a review-count cutoff, or a schema completeness multiplier. Ignore agency posts that invent those numbers.

Perplexity object	How a merchant shows up	What you can actually do
Cited answer	URL in the citation list	Extractable PDP + off-site corroboration
Mention without a link	Brand or SKU named, credit elsewhere	Identity work; see mentions vs citations
Shopping card	Item in their shopping index	Apply / share catalog if you qualify as a merchant partner
Buy-with-Pro / instant buy	Checkout for enrolled merchants	Payment integration they describe; not a ranking formula they published

For the open-web path, the work looks like every other answer-engine PDP: specs in text, one SKU per URL, reviews that are real, and third-party pages that do not contradict you. For the shopping-card path, the work looks like a feed: apply, share the catalog, keep live details fresh. Do not assume Google Merchant Center is a documented Perplexity input. Agencies claim “they read Merchant Center as a fallback.” Perplexity’s own merchant post talks about merchants sharing specs with them. Hedge the Google-feed-as-Perplexity-pipeline story until they document it.

Checklist:

• Split logs: Perplexity answer citations vs Shopping cards
• Confirm Merchant Program status (large-retailer program as they described it — ask, do not assume SMB auto-enrollment)
• If not enrolled, treat Shopping cards as out of scope and fight for citations
• If enrolled, treat freshness like Merchant Center: price and stock are part of the recommendation
• Do not buy sponsored questions and call it organic recommendation — they said those products are distinct

What fails first: feed vs page vs schema fights

The failure I keep seeing is not “we forgot AEO.” It is three systems telling three prices. Google’s share-product-data page calls price and stock mismatch a common cause of synchronization issues. Their recommended mitigation is automatic item updates from the website when the feed and the page disagree. If you disable that and let a nightly feed win, AI answers will recommend an offer the cart cannot honor.

Fight	What the shopper / model sees	Cost	What you do instead
Feed price ≠ PDP price	Recommended at $89, cart is $119	Returned ads, angry tickets, possible disapproval	One price source; auto-updates on
Feed in_stock ≠ PDP sold out	Recommended dead SKU	Trust hit; wasted crawl	Content API or hourly availability
Schema GTIN ≠ feed GTIN	Two products	Split reviews, split eligibility	One identifier sheet
JS-only price	Empty or stale offer in shopping crawls	Merchant listing gaps	Price in initial HTML
Specs in CDN images	Overview cites a competitor table	You “rank” and still lose the comparison	HTML spec table
Review widget blocked to Googlebot	Stars for users, nothing for extraction	Fake richness	Pre-render review summary
Category URL marked up as Product	Garbage rich-result attempts	Search Console errors	Markup PDPs only
Marketplace title is the entity Google knows	Recommendation to Amazon	You funded demand for a reseller	Align names; win the brand query

Procedure when a SKU is missing from AI answers:

1. Search Console: is the PDP indexed and snippet-eligible?
2. Merchant Center diagnostics: approved, pending, or disapproved?
3. Diff HTML vs JSON-LD vs feed on name, price, availability, GTIN, brand.
4. Fetch as Google: are specs in the response HTML?
5. SERP / Overview / AI Mode screenshot for the exact prompt.
6. ChatGPT and Perplexity: shopping card vs citation vs absent — different tickets.
7. Only then write copy.

If step 3 fails, stop the content sprint. You do not have an answer-engine problem. You have a data-integrity problem wearing an AEO hat.

How do you measure product recommendations without fake conversion lifts?

You cannot honestly say “AI Overviews converted 12% more.” Google does not give you a product-recommendation conversion API. Search Console’s generative AI report, where rolled out, is impression-led for AI Overviews and AI Mode. It will not tell you which SKU was named in the sentence. ChatGPT and Perplexity do not drop a GSC-equivalent in your inbox.

Build a frozen panel. Same 25–40 prompts, same device class, weekly. Mix “best X for Y” comparison prompts, brand+model prompts, and replenishment prompts. Log surface, whether your SKU appeared, whether the buy link was yours, and whether you were named, cited, or both.

KPI	Source	Honest use	Dishonest use
Overview / AI Mode impressions	Generative AI performance report (if present)	Directional reach	“We are recommended”
PDP clicks from Web search	Search Console	Traffic you can cash	Proof the Overview named the SKU
Manual recommendation log	Screenshot panel	The actual shopping question	One viral ChatGPT screenshot
Merchant Center approvals	Merchant Center	Feed health	AI visibility score
ChatGPT shopping presence	Partner dashboard if you have one; else manual	Feed enrollment health	Vanity “ChatGPT traffic”
Perplexity citations vs cards	Manual + their merchant dashboard if enrolled	Two columns	One blended percentage
Add-to-cart on those landing URLs	Analytics	Did the click buy	“AI drove $X” without UTM discipline

I have been SEO certified since 2021. The certification does not mint a conversion rate for AI shopping cards. 500+ automations built and 20,000+ hours on agentic systems taught me to distrust a dashboard that cannot name the prompt. 35,000+ hours saved for clients is busywork deleted, not a Merchant Center case study. Do not paste those receipts onto a SKU.

Weekly ritual:

• 25–40 prompts frozen for 90 days
• SKU-level: recommended / cited / mentioned / absent
• Engine column: Google / ChatGPT / Perplexity
• Link target: your PDP vs marketplace vs review site
• Feed diagnostics: disapprovals on those SKUs
• No conversion slide until you have tagged landing URLs and a week of clean analytics

Which pages to fix first for AI search (the catalog version of that question):

Priority	URL type	Why it is first	Defer
1	Top-revenue PDPs that already rank	Highest cost if the recommended offer is wrong	Blog explainers
2	SKUs in comparison prompts you already lose	Unique spec gap is visible in the screenshot	New seasonal SKUs with no demand
3	PDPs with Merchant Center disapprovals	Feed blocks shopping and can starve AI listings	Cosmetic schema warnings
4	Variant parents with no buyable URL	Engines cannot recommend an object that is not for sale	Colorway photography
5	Brand+model queries you own in classic Search but miss in Overviews	Eligibility / extractability, then specs	Link building as the first ticket

If informational CTR on the blog fell, that is the traffic-drop spoke, not this one. Product recommendation work is a catalog problem.

What should you ship in one week?

If you only have a week, do not rebuild the theme. Pick ten money SKUs and make three systems tell the same truth.

Day	Owner	Done looks like
1	SEO + merch	Ten SKUs, ten shopper prompts, baseline screenshots
2	Feed ops	Merchant Center: approved vs disapproved; identifier gaps listed
3	Engineering	JSON-LD vs HTML diff on those ten; price and availability in initial HTML
4	Content	Spec tables live on the five SKUs with the worst unique-spec gaps
5	Content	Who-it’s-for / not-for lists; kill synonym adjectives
6	Reviews / legal	Visible ratings match markup; no coupon-as-author
7	SEO	Re-inspect URLs; re-screenshot the panel; write the ticket list for week two

Skip this week: UCP checkout, ChatGPT partner applications that need legal review you cannot finish, a sitewide FAQ generator, llms.txt as a shopping strategy, and any “AI product description rewriter” that clones the manufacturer.

Do this week	Skip this week
Ten-SKU truth layer	Sitewide AI rewrite
Spec tables in HTML	New category microsites
Feed vs page price/stock	Theme redesign
Identifier sheet	Marketplace expansion
Frozen prompt log	A fabricated conversion model

When is this not worth doing yet?

This work is not worth doing yet if the catalog cannot keep price and stock honest for classic Shopping. AI answers will not rescue a feed Google already disapproves. It is also not worth doing as a substitute for eligibility: if PDPs are noindex, blocked, or snippet-disabled, fix that first.

Situation	Do this instead
Merchant Center is a graveyard of disapprovals	Feed ops, not AEO copy
PDPs not indexed	Crawl/index, then return
Marketplace-only brand with no indexable PDP	You do not have a page to recommend; win the marketplace listing first
Custom / configure-to-order with no stable SKU	Document the configured product as a real URL before you ask engines to recommend it
Category prohibited by OpenAI or Google shopping policies	Stop pitching ChatGPT cards; decide if Search citations are even allowed
Leadership wants a conversion lift slide by Friday	Refuse the slide; offer the prompt log
You sell a service, not a product	This spoke is the wrong tool; use the AEO playbook on the service page

Hire a visibility audit when the feed, the theme, and the answer logs disagree and nobody owns the identifier sheet. DIY the ten-SKU week if you already have Search Console, Merchant Center, and a developer who can put Product in the initial HTML.

Lane work lives on /visibility. The system map is still the playbook. This page is the commerce layer of that system, not a new religion.

FAQ

How do I get products recommended in AI answers and AI Overviews?

Make the SKU a consistent, unique fact: crawlable specs, Product schema that matches the PDP, real reviews attached to a stable identity, and a Merchant Center feed that agrees with the page. Google documents that generative responses can include product listings and that Merchant Center can help; it does not document a special Overview schema. ChatGPT shopping needs an approved product feed. Perplexity Shopping needs their merchant catalog if you want cards, plus ordinary citations if you want answers.

How do I measure whether products are getting recommended in AI answers?

Log a frozen 25–40 prompt panel weekly, SKU by SKU, and record recommended, cited, mentioned, or absent on Google, ChatGPT, and Perplexity as separate columns. Use Search Console’s generative AI report for Overview / AI Mode impressions when you have it, and Merchant Center diagnostics for feed health. Do not treat a shopping-card screenshot as a conversion rate.

What usually fails first when teams try this?

Price, availability, and identifiers disagree across HTML, JSON-LD, and the feed. The second failure is specs that exist only in images or JavaScript, so comparison prompts have nothing to lift. The third is treating ChatGPT and Perplexity like Google and shipping schema tickets for engines that want a partner feed.

How long does this take to show results?

Feed and schema consistency can show up in Merchant Center diagnostics within a crawl or a feed fetch — often days, sometimes longer on slow SKUs. Overview and chat recommendations are non-deterministic; give a frozen panel four weeks before you call a miss. ChatGPT and Perplexity merchant onboarding are partner queues, not a 48-hour content deploy. Anyone selling a date-certain “we will be in AI Overviews” is guessing.

What should I skip if I only have a week?

Skip UCP, sitewide AI rewrites, and llms.txt-as-shopping. Spend the week on ten money SKUs: identifier sheet, feed vs page vs schema diff, spec tables in HTML, and a baseline prompt log. If those ten still contradict themselves, a new protocol will not save the catalog.

When is this not worth doing yet?

Skip it while Merchant Center is disapproved, PDPs are not indexable, or you have no stable SKU URL. It is also the wrong project if leadership only wants a fake conversion lift, or if the category is blocked on the shopping surface you care about. Fix eligibility and data integrity first; then come back to unique specs.

CTA

If the catalog tells three prices, AI answers will pick a fourth — get the SKU honest, then ask to be recommended.

Lane: /visibility · Book a visibility audit.