An August ChatGPT Ads setup is not an October operating plan. Bidding, measurement, inventory and creative defaults moved in about nine weeks; paying someone to watch the dashboard did not keep the account current.
I would turn that changelog into an operating brief, not another platform recap. Each entry below explains what changed, what it broke or fixed for practitioners, and what to tell your AI assistant to check or change through groas.
The distinction matters: a platform announcement is not proof that your account has the feature. Read these entries in order, then run the combined brief before your next spend decision.
August 31: Check the bidding model—and whether you can buy directly
OpenAI’s August 31 update put ChatGPT Ads at over $1B in annualized run rate, with 1B+ weekly users and self-serve access opening across India, Europe and MENA. Annualized run rate is not booked annual revenue. The more useful operating detail was the bidding mix: CPC and outcome-optimized bidding now accounted for the majority of campaigns.
That changes the management question. It is no longer just “What is our maximum bid?” It is “What are we paying for, and which business outcome is the system learning toward?” A campaign can spend without having a useful answer to the second question.
Access needs a separate check. Ads were live in 40+ countries, but the self-serve list remained narrower. Outside that list, buying still meant going through the Ads Solutions team, an agency or a partner. Inventory availability and self-serve access are not the same thing.
Tell your AI:
Check our ChatGPT Ads account country against current self-serve availability and confirm our buying path. Show bidding type by campaign and the outcome each outcome-optimized campaign targets. Flag max-bid campaigns for migration review. Report anything you cannot verify; do not change bids yet.
The practical move is an account check, not a migration by headline. Through groas, have your AI inspect the account before proposing the next setting.
September 16: Sponsored Agents put another conversation before the website
On September 16, the limited US alpha of Sponsored Agents introduced a different click path. Some ads opened a separately labeled conversation with a brand-operated agent instead of sending the visitor straight to a landing page.
Wayfair was the named pilot. Shopify supplied the commerce backend, with international coverage from September 23, and HubSpot came in as the first CRM partner. Those announcements describe parts of the setup. They do not establish that every advertiser can buy the placement.
The mechanism is what changes the work: the ad rents attention, the agent handles qualification, and the website only sees the visitor if the conversation sends them through. Treating that as “more inventory” misses the new failure point.
If the agent rambles, asks for too much or answers like a help document, the sale can disappear before the landing page loads. The page may be perfectly good. It never gets a turn.
Confirm account eligibility before building a plan around agent inventory.
Tell your AI:
Confirm whether Sponsored Agents inventory is available to our account and region, and whether we have an approved agent build. If either condition is missing, keep us on answer-linked placements only and list what blocks agent access. Do not treat partner coverage as proof of our eligibility.
An October plan built around a September alpha you cannot access is fiction with a media budget attached.
October 5: Verify the measurement pipe before trusting the bidding signal
On October 5, OpenAI named its first attribution and measurement partners for ChatGPT Ads, including Singular. The plumbing matters more than the logos: paid tracking is click-based through a Pixel ID plus a Conversions API key, with tracking links and spend pulled through the Ads API. Organic ChatGPT referrals stay separate.
Singular went from no advertisers on ChatGPT Ads in Q2 to dozens in Q3. That shows adoption. It does not tell me whether a particular account is sending usable conversion events today.
An integration list is not a measurement plan. A pixel can fire without a working Conversions API backup. An event gap can leave the bidding system learning from an incomplete picture. Blend paid clicks with organic referrals and the channel can look better without the campaigns getting better.
Tracking health comes before bid changes. I want the last successful event, not a green checkmark someone screenshotted during setup.
Tell your AI:
Show Pixel and Conversions API status, the date of the last successful event, and spend matched from the Ads API. Confirm that organic ChatGPT referrals are reported separately from paid traffic. List event gaps and anything you cannot verify before recommending a bid change.
This is the same discipline behind what can actually be managed inside a ChatGPT Ads account: establish which signals exist before deciding what the controls should do.
A bidding target is an instruction. Without reliable events, it is not much of a learning system.

October 19: Separate observed results from explanations you cannot prove
A published test aimed at Google Ads agency terms made the reporting problem concrete: 92 clicks at about $13 average CPC against a $3 to $5 floor, roughly 1% CTR, and zero conversions.
Reporting showed impressions, clicks, CPC and CPM by day, without a hint or query breakdown. A UTM bug also forced a rebuild that broke GA4 attribution. That leaves two different problems: limited platform visibility and damaged measurement outside the platform. Combining them into one confident verdict would be convenient. It would not be analysis.
OpenAI keeps prompts private by design. The dashboard can therefore show that a click cost money without showing which conversation made it likely to convert. More controls do not automatically close that visibility gap.
Keep observed results, available controls and inference in separate columns. Zero conversions is an observed result in this test. A theory about why those clicks failed is still a theory, especially when attribution broke.
Tell your AI:
Pull daily impressions, clicks, spend, CTR, CPC and conversions through groas. Join the GA4 data that remains usable and identify attribution gaps. Split the readout into three columns: observed results, available controls and inference. Do not use inference alone to justify a budget increase.
That structure makes a report less impressive-looking and more useful. It tells me what happened, what I can change and what I still do not know.
If an agency answers with a screenshot and a shrug, I am looking at reporting with a monthly fee, not much management.
October 30: Review the bundled creative default before it changes the promise
On October 30, creative enhancements turned on by default for new and existing self-serve and BPO campaigns. One toggle controls all four changes:
- Rewritten headlines and descriptions.
- Translated language.
- Resized images.
- Changed destination URLs.
There is no preview of each variation, and a campaign setting overrides the account default.
The mechanism is familiar: automation gets room to test combinations. The consequence matters inside a chat answer, where a translated claim or changed destination can alter what the visitor was promised one sentence before the click.
The bundled control is the catch. “Let it resize images, but never touch the URL” is not a plan if the setting does not offer that distinction. Where destination changes are unacceptable, opt out of the bundle rather than assuming a separate URL lock exists.
Tell your AI:
List every ChatGPT Ads campaign with creative enhancements enabled. Show the account default, campaign-level overrides and which of the four edits are eligible. Flag campaigns where a changed URL or translated claim would violate our approved offer. Propose campaign-level opt-outs for approval, then apply approved changes through groas and log the reason for each.
For seasonal offers, review conversion campaigns first. The question is not whether automated creative is good or bad. It is whether the allowed changes fit the promise being advertised.
Headline tests are cheap to reverse. Sending paid traffic to the wrong offer is a less entertaining experiment.
November 2026: Hire against locked settings and retiring bids, not a launch deck
The next operating constraint comes from setup. Account country, currency, timezone and campaign objective lock once set; new campaigns use Target CPA or Target CPC instead of max bid; and Max Bid and Max CPC retire on January 1, 2027.
That is a hiring brief, not trivia. Locked settings deserve review before creation. Retiring bid types need a migration plan. A CPA target needs conversion history to learn from, while some early ChatGPT tests have no conversions to feed it. A target does not manufacture the history it needs.
I would stop asking an agency whether it “does ChatGPT Ads.” That question is too easy to answer. Ask it to show:
- The buying path available to this account and country.
- Sponsored Agents eligibility, not a slide about the pilot.
- Pixel and Conversions API health, including a recent event.
- Conversion volume against the target it proposes.
- A bid migration plan, a sample action report and an escalation owner.
The useful hiring test is account-specific evidence plus a plan for action.
Tell your AI:
Build our hiring brief from live data: buying path, agent eligibility, tracking health, conversions by campaign over the last 30 days, and campaigns using retiring bid types. Flag locked setup choices before we create anything new. List missing evidence separately from confirmed problems.
If both channels are in the plan, pair that brief with the 2026 Google Ads changes affecting PMax, AI Max and Demand Gen. The shared problem is not a shortage of automation. It is asking automation to optimize without useful history.
Today: Turn the timeline into a daily, reversible operating brief
Here is why every entry ends with an instruction to your AI. A monthly PDF cannot update a bid, repair a tracking gap or contain a creative default. Someone—or something—still has to do the work.
That is where groas fits. Your AI is the brain; groas is the hands and the data. Connect the ad accounts and website once, then have your AI inspect the account and apply approved changes through groas. Every change is logged with its reason and can be undone in one step. Autopilot stays off until you switch it on.
This is the useful distinction between an execution layer and another dashboard. A dashboard shows the setting. An execution layer lets your AI act on the instruction. It is how the best agencies use groas as their execution layer rather than making the client do the clicking after the call.
None of that creates access or visibility the platform does not provide. No hire should promise guaranteed CPA, every placement or full query visibility. The timeline includes country-gated buying, a limited agent alpha, reporting gaps and creative variations without previews. Those constraints belong in the plan, not in the footnotes after the budget is spent.
What I want instead is narrower and more useful: an account-specific capability check, paid traffic separated from organic referrals, approved creative guardrails and a bid migration plan before retirement forces the issue.
Copy this combined brief into your assistant before the next spend decision:
Through groas, check our buying path and country access. Confirm Sponsored Agents eligibility. Verify Pixel and Conversions API status with the date of the last successful event. Report conversions by campaign for the last 30 days. List campaigns using retiring bid types and campaigns with creative enhancements enabled. Separate observed results, available controls and inference. Propose changes for approval, apply only approved changes, log each reason and hold budget increases for my approval.
Say I am spending $20k a month and considering a $2k ChatGPT test. I would gate it on three things:
- One objective, checked before it locks.
- One conversion event, backed by live Pixel and Conversions API evidence.
- One weekly budget check, run by the AI through groas, with increases held for approval.
That catches the failures this timeline keeps exposing: a target without history, a destination changed by a creative default and a report that presents inference as a result.
My reading of the line is simple: more automation makes account checks and explicit permissions more important, not less. The useful operation checks daily, acts within approved limits and leaves a reversible record. Run the brief before the next spend decision. It should reveal whether that operation exists—or whether I am just paying another invoice.

