Chris Shuptrine · chris@getchatads.com
What //Commerce sees today. A reader lands on a publisher's article. The page renders with anchor tags. //Commerce rewrites the outbound link. Sovrn earns. Publisher earns. Everyone wins.
What's happening now. An AI Overview, ChatGPT answer, or Perplexity response uses the same publisher's content to produce generated text with the same product mentions. No anchor tag. No page render. No URL the publisher authored.
//Commerce's pipeline never runs. Not because it's slow or broken — because the input it was built for doesn't exist on this surface.
The deck below is one engine that closes this gap, and the math on what it would cost to build it inside Sovrn versus integrate it from a working v3.
This is the production endpoint, hit live. The intent gate decides whether the sentence is monetizable, the retriever picks the SKU given context, the rewriter returns a tracked link.
Why this is the slide that matters: nothing in this trace exists in //Commerce. There is no retriever to call. There is no intent gate. The catalog is keyed on URLs that aren't present in the input. Building this from scratch is what the next-but-one slide is about.
ChatAside running on a real customer (The Copper Pot, 98+ recipes). Same publisher Sovrn would sell to. The widget reads the page, answers in context, and routes monetization through ChatAds. One JS snippet to install.
A four-to-six hour cold smoke over alder gives you the silky texture you want without ever crossing into "cooked." The temperature management is the whole game — too warm and you lose the gloss…
…hold a steady internal temp between 75°F and 90°F throughout the smoke. A good leave-in thermometer is non-negotiable here. You're not opening the lid to check…
…the cure is the easy part. The brine takes about 18 hours at 38°F in the fridge…
Why this is the slide that matters for Sovrn: this is the publisher relationship Sovrn already has — same install motion (one script tag), same rev-share model, same customer. The widget is the new on-page surface that monetizes the conversations //Commerce can't see.
model2vec + MLP, native Go. Decides whether a given AI sentence should be monetized at all. Suppresses "talk to a doctor," "grab one at any store."
Latency: <1ms. The unsexy core. //Commerce has no analog — HTML doesn't need one.
e5-small-v2 ONNX encoder + FAISS ANN over a 1M-SKU index. Embeds query + surrounding context together — context is what disambiguates "meat thermometer for salmon" from "meat thermometer for steak."
17ms encode + 3ms search = 20ms. Scales sub-linearly to 10M+ SKUs.
Takes the SKU + the surface context (which publisher, which API key, which network) and emits the tracked link. Pluggable per network — Amazon today, anything tomorrow.
Latency: ~6ms. The layer where Sovrn's merchant graph would slot in.
What this isn't. It isn't a wrapper around GPT-4 or Claude. There is no LLM round-trip on the request path — an LLM call alone is 1–2 seconds, which busts the budget before you even start. The whole stack is deterministic NLP plus a sidecar embedding service. That's the engineering shape that has to exist for AI commerce, and it isn't the shape //Commerce has.
An honest engineering inventory of what would need to ship before //Commerce can do what the trace on slide 3 does. None of this is novel research — but all of it is real work that doesn't share a codebase with the URL-rewriting pipeline.
| Component | What it actually entails | Eng months | ChatAds |
|---|---|---|---|
| Intent classifier | Label corpus of AI-generated sentences (monetizable vs. not), train + serve a model under 1ms. No prior art at Sovrn — HTML never needed one. | 2–3 | ● |
| Embedding pipeline | Pick an encoder (or train one), set up ONNX serving, get encode latency under 20ms at production QPS. | 1–2 | ● |
| Catalog indexing | Re-shape the merchant graph into an SKU-level vector index. Backfill, incremental updates, sharding past 5M SKUs. | 2–3 | ● (Amazon only) |
| Hot path runtime | Go service with sidecar architecture, <100ms budget end-to-end. Not retrofittable onto an async HTML-mining pipeline. | 2 | ● |
| Scoring harness + eval corpus | The thing that tells you whether you're at 60% or 89%. 100+ evaluated runs takes months to accumulate. Without it you can't ship safely. | 2 | ● (89% today) |
| Publisher widget | Embeddable JS, page-context Q&A, hosted LLM inference, content training, install / dashboard. Same shape as a //Commerce install — different code. | 2–3 | ● |
| SDKs & integrations | npm + PyPI + MCP + n8n + Zapier + Lovable + Replit + ChatGPT Apps + Claude Code. The surface area for the AI-builder customer. | 2 | ● |
| Total | Plus integration, hardening, the inevitable scope expansion. | ~12 months · 2 ML hires |
The strategic risk isn't cost. It's time. A year of build is a year of competitors moving into AI commerce while //Commerce sits on the sidelines. The retriever, the intent gate, the widget, the eval corpus, and the integration surface already exist at v3. The question for Sovrn product is whether bolt-on or rebuild is the right path — not whether to ignore it.
Not a six-quarter roadmap. A concrete bring-up where each phase has a thing that goes live.
Nothing in this deck is a commercial proposal. The point is to put the engineering and product folks on both sides in front of the same diagram and argue about the right things.
Chris Shuptrine
Product: getchatads.com · chataside.com
Docs: docs.getchatads.com