The current turn is not the whole buy.
On the Mobile Dev Memo podcast with Eric Seufert (September 30, 2026), Asad Awan, who leads monetization product and engineering at OpenAI, described how ChatGPT Ads selection can lean on what the product already remembers about a user across threads, not only on the words sitting on screen. The Keyword’s October 1 write-up put the operator sentence in plain English: ads chosen partly by memory, not only by the conversation (The Keyword, Kole Ogundipe, October 1, 2026; Mobile Dev Memo, September 30, 2026).
That is not a cute personalization footnote. It is a media model change. Keyword systems map “Yosemite trip” to a hotel or a flight. Memory can surface hiking gear for a user already known to hike, with explainer text above the ad saying why it appeared. Awan called that “contextual relevance but personalized based on a little bit longer history.” A second, purely personalized format can sit in tool-like moments (he named the wait while an image generates) and compete with contextual ads for the same placement (Mobile Dev Memo transcript; The Keyword, October 1, 2026).
If your ChatGPT Ads brief still reads like a search keyword sheet tied to the live turn, you are buying yesterday’s surface. The control surface that matters now is what you feed the system about the product, who it is for, why someone should care, and how privacy and consent fence personalization.
What Awan actually said
Keep the claim exact.
Users run multiple journeys at once across conversations. ChatGPT’s memory product bridges preferences across threads, so a seemingly unrelated new chat can still carry learned preferences. That bridge already shapes organic answers. Awan said the same bridge enables better ads: intentional and prospecting, discovery and late-journey delivery (Mobile Dev Memo, September 30, 2026).
He gave three classes of work beyond picking the SKU. Dynamic creative explains the match (a nurse on their feet for long hours sees cushioning language on a sneaker). Variant selection uses product feeds for size and color, not only product. Post-click, Sponsored Agents put the user into a clearly labeled conversation with the brand while the original ChatGPT thread stays separate (Mobile Dev Memo; The Keyword, October 1, 2026).
He also drew hard inventory limits. “Not every conversation is inventory.” A photo of a child’s bruise is the kind of conversation that should never carry an ad. Memory also works against repetition: after a hotel booking, stop showing hotels (Mobile Dev Memo; The Keyword, October 1, 2026). OpenAI’s Help Center already excludes ads near sensitive or regulated topics including personal health, mental health, or politics, and says Temporary Chats will not show ads (OpenAI Help Center, Ads in ChatGPT, accessed October 2, 2026).
Scale claims on the record from Awan: ChatGPT Ads in more than 40 countries, tens of thousands of advertisers, and a $1 billion annualized revenue run rate in under 200 days. Roadmap priorities he named: more countries, measurement that fits advertisers’ existing tools, and format work on Sponsored Agents and personalized ads across verticals (Mobile Dev Memo; The Keyword, October 1, 2026).
Trust rubric, Help Center, and the personalization toggle
Awan’s internal rubric for every ad decision: user trust outranks user value, user value outranks advertiser value, advertiser value outranks revenue. The ad system optimizes an “ad score” that combines user value and advertiser value. Core-model answers stay independent of ads. OpenAI’s published principles rule out optimizing ChatGPT for time spent. Paid plans let users opt out of ads (Mobile Dev Memo; OpenAI Help Center; OpenAI advertising principles page linked from Help).
Help Center is the operator source of truth for what personalization can touch. Ads may appear for Free and Go plans. Plus, Pro, Business, Enterprise, and Edu do not have ads. Ads do not influence answers. Selection can consider current conversation context and intent, landing page, title, copy, advertiser context hints and targeting selections, and, when ads personalization is enabled, select signals from the broader ChatGPT experience. When personalization is on and memory is on, ChatGPT may save and use memories and reference recent chats when selecting an ad. Users can turn off ad personalization, clear ads data, and manage memory. If personalization is off, ads still use the current thread but not other threads, ads history, or topics for selection. Personalized ads are not initially available in the EEA or Switzerland (OpenAI Help Center, Ads in ChatGPT, accessed October 2, 2026).
Advertisers do not receive chats, chat history, memories, or personal details. They get aggregated performance such as views and clicks. That privacy fence is load-bearing for any brand that wants memory-assisted relevance without owning the user’s history (OpenAI Help Center).
Context hints and product philosophy are the buy
Awan’s answer to what advertisers should supply was blunt. The most useful input is what the business offers: products, landing pages, product feeds, and the “philosophy” behind a product. Context hints let advertisers describe brand voice and who a product is for. He framed that as telling the system “why should somebody care about this,” not as classic targeting. Conversion data from Pixel and Conversions API still matter as a check on whether stated preference becomes action. Buying is moving toward automation: autobidder alongside max bids, Shopify and HubSpot integrations, prompt-based ad creation through a ChatGPT Work plugin (Mobile Dev Memo; The Keyword, October 1, 2026).
OpenAI’s own Context Hints help article matches that story. Hints live at the ad group level. Strong hints are additional, specific, relevant, and accurate. They are not targeting rules or delivery instructions. “Residential plumbing services available in Chicago” describes the business. “Show this ad only to people in Chicago” is a delivery instruction, and hints cannot enforce that. Write natural phrases about what, who, and when, not disconnected keyword lists (OpenAI Help Center, Write Context Hints for ChatGPT Ads, accessed October 2, 2026).
Read that next to Awan’s memory thesis and the media consequence is obvious. When selection can pull from longer history, and when dynamic creative and variant selection assemble the unit, your feed accuracy and hint quality become the creative brief. A thin catalog and vague “lifestyle brand” hint will lose to a competitor that states concrete use cases, constraints, and philosophy the ranking system can match to remembered preferences.
The Keyword’s framing is right for growth teams: what brands feed into ChatGPT increasingly becomes the ad itself (The Keyword, October 1, 2026).
What operators will get wrong
Briefing keywords for the live turn only. You will optimize for Yosemite-equals-hotel and miss hiking-gear-from-memory competition for the same slot.
Treating context hints as keyword stuffing. Help Center says natural phrases, one idea, not delivery instructions (OpenAI Help Center, Context Hints).
Shipping a product feed that cannot support variant selection. If size and color are garbage, Awan’s second class of work fails before creative starts (Mobile Dev Memo).
Ignoring the personalization and memory toggles in creative and measurement plans. Help Center gives users an off switch. Regions without personalized ads at launch (EEA, Switzerland initially) need a contextual-only plan (OpenAI Help Center).
Assuming every chat is inventory. Awan and Help Center both fence sensitive conversations. Brands that push into health-adjacent journeys without reading the exclusion policy will burn trust and waste ops time (Mobile Dev Memo; OpenAI Help Center).
Confusing “philosophy” with brand slogan. Awan’s ramen example was about why the product exists and who should care. That is briefable product truth, not a tagline dump (Mobile Dev Memo).
Skipping Pixel/CAPI grounding because “users tell ChatGPT what they want.” Awan still called conversion data the check on stated versus revealed preference (Mobile Dev Memo).
Who this is for this week
Good fit: performance and brand teams already live in ChatGPT Ads Manager or API partners, with clean product feeds, landing pages that state real use cases, and legal comfort with Free/Go inventory plus consent controls. Commerce brands that can write concrete who/when hints. Agencies that can rewrite ChatGPT briefs away from keyword maps toward philosophy, feed QA, and privacy fencing.
Poor fit this week: teams that only want a search-style keyword buy glued to the current prompt; brands without feed hygiene; anyone treating memory personalization as always-on globally without checking Help Center regional limits; orgs that cannot staff Sponsored Agent handoff rules when post-click chat is in scope.
What to do this week
Audit every live ChatGPT Ads ad group for context hints. Replace keyword piles with one-idea phrases that state what, who, and when. Keep delivery controls in the product’s real targeting tools, not in hint text (OpenAI Help Center, Context Hints).
Rewrite the product philosophy block for your top three SKUs. One paragraph each: why it exists, who it is for, what problem it solves better than the obvious category substitute. That is the material Awan said the system needs (Mobile Dev Memo).
Run a feed QA pass for variant fields the ranking system can use (size, color, availability, key attributes). Broken variants are broken personalization.
Map privacy and consent for your markets. Document Free/Go vs paid plan inventory, personalization toggle behavior, EEA/Switzerland limits, Temporary Chat exclusion, and sensitive-topic fences. Put that map in the media brief before anyone celebrates “memory targeting” (OpenAI Help Center).
Build a two-lane creative test plan: contextual-only versus contextual-plus-personalization where personalization is available. Track whether memory-assisted relevance reduces wasted hotel-style category matches on journey chats, without inventing attribution theater.
Decide Sponsored Agent policy now. Clear brand labeling, handoff from organic thread, and what the agent is allowed to say. That is Awan’s third class of work, and it sits next to memory selection whether you are ready or not (Mobile Dev Memo; The Keyword).
Rebrief media and brand together. Media owns placement and bidding. Brand owns philosophy and claim accuracy. Catalog owns feed truth. Privacy owns consent fencing. Memory makes those four seats share one control surface.
Awan’s own hierarchy still applies: trust, then user value, then advertiser value, then revenue. The memory is not the conversation. The conversation is one signal. The longer history is another. Your job this week is to feed the system product truth it can match to that history, inside the fences OpenAI already published, instead of buying keywords for a single turn that already forgot what the user said last week.
