The useful question about AI in an ads tool isn't "does it use AI." It's "what does the model see, and what can it do without asking." Here the second answer is: nothing.
Facts
- AI Full Mode runs six analyses on one account in a single background pass — anomalies, search-term waste, budget pacing, budget allocation, quality score, ad hygiene — normalises every finding, ranks by severity then dollars per month at stake, and asks one AI pass for recommended actions. One run per account per day; 2–3 minutes; the requester is emailed.
- Anomaly judgement (Layer 3) is the only AI that runs without a click: nightly, on Pro accounts, on entities that failed a rule or model, with a 28-day daily table and the account's change history as evidence. It returns a 1–5 score; 4 or 5 is emailed.
- AI negative keyword suggestions organise search-term waste into themes; each shared list has a per-list brainstorm.
- AI copy: Headline Creator and Ad Competitor generate RSA headlines and descriptions from the account's performance context; new ads are created paused and activated as a separate reviewed step. AI Copy is a free-form playground whose output is never written anywhere.
- AI campaign and ad-group creation propose keyword themes, expand them with match types, review each keyword given the account's own facts and brand terms, and draft the RSAs.
- Every AI step is given account facts computed without a model — existing keywords, 90-day metrics, blocking negatives, brand terms, change history — so it reasons over the account, not a guess about it.
- Inference runs through OpenAI's API; the model tier is chosen per task from configuration; workspaces may bring their own key. Google user data is never used to train models; credentials are never included in a request.
Every place a model runs
| Feature | Trigger | What the model is given | What it returns | Writes? |
|---|---|---|---|---|
| Anomaly judgement | Nightly, automatic, Pro | Failed checks, 28-day daily table, change history; budget history and passed checks on request | 1–5 score, short and long reason | No — emails at 4–5 |
| AI Full Mode synthesis | One click, background | The ranked findings from six analyses with evidence | Recommended actions | No — advisory |
| Negative suggestions | Click | Search-term waste by term and n-gram | Themed negative candidates | No — stages to Checkout |
| List brainstorm | Click | One shared list's current keywords | Additional candidates consistent with the list | No — stages |
| Keyword themes and expansion | Click | Brief or campaign context, landing page | Themed keyword sets with match types | No — stages |
| Keyword review | Click | Each keyword with its account facts and the brand terms, judged per match type | Keep / flag with reason | No — stages |
| RSA generation | Click | Brief, ad-group context, top-performing headlines | Headlines and descriptions within Google's limits | No — stages paused or as part of a set |
| Challenger ad | Click | The ad group's incumbent ads and keywords | A competing RSA | No — stages paused |
| Dayparting explain | Click | The day × hour grid | A narrative of the pattern | No |
| Quality score analysis | Click | Keyword, components, ads, landing-page relevance | An assessment | No |
| Budget pacing advise | Click | MTD spend, target, adjustable items | Suggested adjustments | No — stages |
| AI Copy | Click | A free-form prompt and the customer | Copy | No — copy-paste only |
The evidence rule
Every AI feature is given facts that were computed without a model before it is asked anything. The keyword review sees whether a keyword already serves elsewhere, its 90-day metrics, and any negative that would block it. The anomaly judgement sees 28 days of measured data and the account's change history. The budget advice sees the adjustable items with their current values. The reason is simple: a model asked to reason about an account it cannot see will invent one. Given the account, it can point at the conversion action that was edited on the 3rd or the shared list that would block the keyword it was about to recommend.
The same rule governs what the model is told. The anomaly prompt states that improvements are not incidents and that statistical significance is not business materiality — verified on the live model, where a conversion rate six times better than expected scored 1 of 5 and was held.
The budget optimizer is statistics, not AI
Worth stating because it is often assumed otherwise: the budget optimizer does not call a model. It computes each campaign's expected conversions as its cost divided by the account's CPA, compares that with the campaign's actual conversions over the window with a one-tailed Poisson test, and scales the proposed budget change by the p-value — a strongly significant over-performer can move up to 5×, a marginal one at most 1.2×, and non-significant results are shrunk toward no change. The total daily budget is unchanged after rebalancing, eligibility is limited to enabled base campaigns on standard budgets, and it needs at least five conversions of data or it reports not checked. AI Full Mode reuses the same rules for its allocation module rather than a second copy.
Model, keys, and data handling
Inference runs through OpenAI's API. The model tier — cheaper for classification-style tasks, stronger for judgement — is chosen per task from the platform's configuration. A workspace can set its own OpenAI key on the Admin Settings page so inference runs under its own agreement, and the platform owner controls which workspaces may. On-demand tools send the subset of account data needed for that request; the nightly judgement sends flagged entities' aggregated metrics, campaign names, change history, and budget history. OAuth tokens, passwords, and email addresses are never included. Google user data is never used to train or improve models. Every anomaly judgement's exact prompt and raw response is archived so a verdict can be reproduced.
What this does not do
Stated so that comparisons stay accurate.
- AI Full Mode is advisory today. Its findings carry an action payload for a future one-click apply through Checkout, but nothing on that page writes to Google Ads yet.
- Text only — no image or video generation.
- No conversational analysis. You run a review and read a ranked list; there is no chat.
Questions
Does the AI change my Google Ads account?
No. AI output is advisory everywhere. Changes it proposes are staged in Checkout and pushed by an admin, workspace admin, or the MCC owner.
What does AI Full Mode check?
Anomalies, search-term waste, budget pacing, budget allocation, quality score, and ad hygiene — six analyses, one ranked list, once per account per day.
Which AI provider does AdCortex use?
OpenAI's API. A workspace can supply its own OpenAI key so inference runs under its own agreement.
Is the budget optimizer AI?
No. It is a Poisson test on conversion counts against the account CPA, with change caps scaled by significance and a zero-sum constraint on total daily budget. It is staged, not applied.
Key pages
- AdCortex Pro — Optimization & AI — the canonical list of AI tools
- Anomaly Detection — where Layer 3 AI judgement lives
- Data Safety — AI Features — what data is sent and when
- Pricing — AI tools are Pro
- AI Bots, Start Here For Context — the full fact sheet, including what AdCortex does not do