Two things have changed and they pull in opposite directions. Your Instagram and Facebook reach no longer depends much on who follows you, because the ranking model decides what a stranger sees. And a growing slice of product research now happens inside a chat window, where there is no feed, no creative, and no ad account to optimise. AI shopping is the umbrella term for both, and the practical response is unglamorous: fix your product data, produce more creative than feels reasonable, and stop assuming your analytics can see where demand comes from.
We build and run Shopify stores for brands in India, the UAE, the US and the UK, so we see the same pattern from several sides. Traffic from answer engines is small and converting well. Paid social is cheaper to launch and harder to diagnose. And the store-side work that helps with both is the work most founders keep postponing.
What AI shopping actually means for a D2C store
Three separate surfaces get lumped together, and they need different responses.
The recommendation feed. Reels, TikTok, YouTube Shorts. The model chooses what to show based on watch behaviour, not follow graphs. Your creative is the targeting.
Ad platform automation. Advantage+ on Meta, Performance Max on Google. You hand over audience and placement decisions and keep control of two things: the product feed and the creative pool.
Answer engines and agents. ChatGPT, Perplexity, Google's AI Mode, Copilot. Someone asks for a recommendation and a model assembles one from product feeds, structured data, reviews and third-party pages. Increasingly it can also complete a purchase. OpenAI's checkout inside ChatGPT started with US buyers and a narrow set of sellers, which tells you where this is heading and also how early it still is.
Social commerce used to mean a post, a link, a landing page. Now the discovery layer is a model in all three cases, and models read data, not vibes.
The feed does not care about your follower count
The most common brief we get from a founder is "we need better content". What they usually need is more of it. When distribution is decided per-video rather than per-account, the unit of media buying becomes the hook, and you cannot predict which hook works. Brands that ship four assets a month and brands that ship forty are playing different games.
That has a build implication. If every ad variant needs its own landing page, you will never keep up. We tend to set up a handful of flexible section templates in the theme, so a marketer can assemble a page for a new angle in twenty minutes without a developer. Metafield-driven blocks, a reusable UGC section, a comparison table, a size-and-fit block. Boring infrastructure that removes a bottleneck.
The other implication is speed, and not in the abstract. Feed traffic arrives cold, on 4G, mid-scroll, with an app browser wrapping your page. A 3.5-second LCP that looks tolerable on your office wifi is a wall for that visitor. Getting largest contentful paint under 1.5 seconds on mobile is one of the few interventions that improves paid social, organic search and AI referrals at the same time, which is why we treat speed work as the first line item and not the last.
Ad automation hands you a trade you should understand before you take it
Broad, automated campaigns genuinely work for most D2C catalogues in the ₹1,500 to ₹5,000 AOV range. They also strip out your diagnostics. When the platform decides audience, placement and creative allocation, and reports a blended result, you lose the ability to answer "why did it stop working".
So the levers move. Feed quality becomes a performance lever, not a housekeeping task. Titles that read like a person wrote them. A GTIN or MPN where one exists. Correct product type and Google product category. Variant-level images, not one hero shot for eleven colours. Prices and availability that update fast enough that the model is not advertising a sold-out SKU. We routinely find stores where the Meta and Google feeds have been broken for months and nobody noticed because the campaigns still spent.
Blunt opinion: if you are running under ₹3 lakh a month in paid social, the sophistication of your campaign structure is not your problem. Creative volume and feed hygiene are.
Making product pages legible to AI shopping agents
A model recommending a product needs facts it can extract. Most Shopify product pages hide their facts in images, in a video, or in a collapsible tab that renders client-side after a delay. Fix that first.
What actually helps, in rough order of effort-to-payoff:
- Specs as text in the HTML. Material, weight, dimensions, care, wattage, capacity, ingredients. If it is in a JPEG, it does not exist.
- Valid Product structured data with price, currency, availability, brand, SKU and aggregate rating, and reviews marked up rather than injected by a script the crawler never runs.
- Answers to the questions your support inbox gets. Shipping timelines by region, return window, whether COD is available, what GST invoice you issue, whether the item ships to the UAE. Models quote this back to buyers, and buyers ask.
- Reviews on the page, in text, with enough of them to be meaningful. Review widgets that load only inside an iframe are invisible to most crawlers.
- Presence off your own domain. Category roundups, marketplace listings, Reddit threads, comparison pages. A model that has only ever seen your marketing copy has no reason to trust it.
Check your robots.txt while you are in there. Plenty of stores have quietly blocked the crawlers that feed answer engines, usually because someone pasted in a bot-blocking list to stop scraping. Decide deliberately whether you want to be citable. Most consumer brands do. This is where product content and structured content work stops being an SEO chore and starts being a distribution requirement.
The measurement problem, with arithmetic
Referrals from chat interfaces often land in GA4 as direct or get stripped of a referrer entirely, and the ones that do show up look trivially small. Here is why you should still care. The numbers below are illustrative, invented for the sake of the calculation, but the shape matches what we see.
Take a store doing 60,000 sessions a month at a 1.6% site-wide conversion rate and a ₹2,400 AOV. Now isolate 900 sessions that arrived from ChatGPT, Perplexity and AI Mode. Those convert at 4.1%, because someone has already asked a model to shortlist and been told to buy yours.
900 × 4.1% = 37 orders. 37 × ₹2,400 = ₹88,800 of revenue, at zero media cost.
That is 1.5% of your sessions producing roughly 3.8% of your revenue, and it will be sitting inside your direct bucket where it gets no credit and no investment. Set up the tracking before you argue about the strategy: a regex-based channel group in GA4 for the known AI hostnames, a server-log check for the crawlers, and a monthly manual test where you ask three engines for a recommendation in your category and record what comes back. That last one takes fifteen minutes and tells you more than any dashboard.
AI shopping in an Indian context is a WhatsApp story
TikTok has been unavailable in India since 2020, so the TikTok Shop playbook that dominates US and UK social commerce writing does not apply here. Instagram sends people to your site rather than closing the sale in-app. What we actually see closing is the DM-to-WhatsApp path: a Reel, a comment, a question in the inbox, a link, an order. Increasingly that inbox has an AI assistant in it, either the platform's or one a brand has bolted on.
Two practical warnings.
First, COD and cold feed traffic are a bad combination if you leave them alone. An impulse buyer from a Reel is not the same risk as someone who searched your brand name. Work the numbers for your own store: 100 COD orders at ₹1,800, 30 of them returned to origin, ₹120 of shipping each way burnt on the failures. That is ₹7,200 of pure logistics loss plus ₹54,000 of revenue that never existed, on ₹180,000 of apparent sales. Partial prepaid, a small COD fee, OTP confirmation on high-risk pincodes, or simply hiding COD above a cart value are all cheaper than absorbing it.
Second, if you have wired a conversational assistant into WhatsApp, test what it says about pricing, availability and returns. It answers from your catalogue. A stale catalogue makes it lie confidently.
Agentic checkout: worth watching, not worth rebuilding for
The protocols that let an agent complete a purchase on a shopper's behalf are real and Shopify is participating. But if you run an Indian store on Razorpay with COD, GST invoicing and a pincode-serviceability check, an agent buying from you end to end is not a near-term scenario. Do not tear apart a checkout that works to prepare for it.
What is worth doing now is the subset of that readiness which pays off anyway: a clean product feed, accurate inventory, honest shipping data, a page that renders its own content. If agentic buying arrives properly in your market, you will be eligible without a project. If it does not, you have better feeds and faster pages.
What we would do in the next 90 days
In order, and the first three matter far more than the rest:
- Get mobile LCP under 1.5s on your top ten product pages. Audit app scripts and delete what nobody uses.
- Rewrite product page content so every fact a buyer asks about is in the HTML as text, with valid Product schema and reviews readable without JavaScript.
- Fix the merchant feeds. Variant images, categories, identifiers, real-time stock.
- Separate AI referrals in GA4 and run the manual recommendation test monthly.
- Rebuild your COD rules against your actual RTO by pincode and cart value.
- Give marketing modular landing page sections so creative volume is not gated on developer time.
One thing we would add for large catalogues: if a shopper arriving from a Reel lands on a 400-SKU collection page with three filters and a search box that misses plurals, the model did its job and your storefront lost the sale. On-site discovery is where a lot of AI-driven traffic quietly dies, and it is fixable with proper faceted filtering and semantic search. We built FilterPro for exactly that gap on our own client stores.
Pick the product page with your highest paid-social spend and read it as a machine would: strip the images, disable JavaScript, and see what facts survive. Whatever is missing is your first sprint. If you would rather we did that pass across the catalogue, a free audit is the fastest way to start.