Jev for Google Ads: 8 jobs a decision model can take over

What is Jev: a 45-second explainer for marketers

ChatGPT answers a question by writing a paragraph. Jev answers by picking from a list. Ask both "is this invoice fraud?" ChatGPT types out a sentence in 8 seconds. Jev gets three options, Fraud / Clean / Review, and returns a score for each in 0.1 seconds: Clean, 88%. No writing, no explaining, just the pick. That is Jev: a model that only decides, 100x faster and far cheaper. Every job on this page is a pick from a list.

Jev on 724 live ads in 40 seconds

Matthew Berman ran Jev on 724 live ads from 37 brands. In 40 seconds it labelled every hook, format, offer, CTA and awareness stage, and flagged the landing pages that did not match the ad, for about 9 cents of tokens. This is the job most ad teams do by hand once a quarter, if at all.

What Jev looks like on an ad account: demo

One recording on a sample account. Jev pulls competitor ads, scores every one, and the winners go to an LLM to write briefs. The counters on screen are examples, not measurements.

Meta Ads: find what competitors keep running, then write ads like it

Jev reads 1,600 competitor ads, works out which styles keep running, and scores new briefs in those styles. Whole run: about $0.71.

• Pull 1,600 competitor ads from the Meta Ad Library.
• Tag each ad by type (talking head, myth-buster, price anchor) and check how long it stayed live. Ads that survive 60 days are the ones making money.
• Write and score. An LLM drafts new briefs in the winning styles; Jev scores each on hook, brand fit and how likely it is to survive.

The 8 jobs at a glance

Google Ads work splits into two piles. One pile is judgment calls with a known answer set: is this search term a buyer, does this keyword belong in this ad group, is this headline relevant to this landing page. The other is writing and diagnosis. Jev only fits the first pile, and the first pile is most of the hours.

Every row has a known answer set, which is the test for whether a decision model fits. If the answer has to be written, it belongs to an LLM.

What Jev is, in one table

Jev is a model from TypeSafe AI, the company founded by Diogo Almeida after his work on ChatGPT at OpenAI. It opened in early access on September 15, 2026. TypeSafe calls it a System One model: you send it a block of text (the state) and a set of typed questions, and it returns typed answers with a probability attached. It does not write sentences.

Two caveats belong next to those numbers. The benchmarks are vendor-reported and measure agreement with frontier models, and no large independent reproduction exists yet. And TypeSafe says itself that it cannot prove the launch price is unsubsidized. Treat the price as today's price.

The 8 jobs, one by one

1. Search-term triage

Question: "Is this query from a buyer, a researcher, a job seeker, a competitor search or junk?" Answer type: Choice. Next: Auto-negative above the threshold, approval queue in the middle band.

A 30-day search terms report on a mid-size account is 5,000 to 50,000 rows. Today an LLM reads them in chunks and misses the long tail. One Choice per row covers the whole report every night, and the rule from the account stays in code: never negate a term that converted in the last 90 days.

2. Negative keyword conflicts

Question: "Does this negative block a keyword we are bidding on?" Answer type: Noul. Next: Conflicts go to the approval queue with both terms side by side.

Code pairs every negative with the keywords it could match. Jev answers whether the pair is a real conflict in meaning, which string matching alone gets wrong on close variants.

3. Keyword to ad group fit

Question: "Which of these ad groups does this keyword belong in?" Answer type: Choice. Next: Misplaced keywords are listed with the better home.

Ad groups drift as accounts grow. A Choice across the account's ad groups, run per keyword, produces the restructure list that usually takes an audit to find.

4. RSA asset relevance

Question: "Does this headline match what this landing page offers?" Answer type: Score. Next: Low scores are queued for a rewrite by the LLM.

Every headline and description against its final URL. The score is the cheap half, and the rewrite of the bottom 10% is the half that needs a writing model.

5. Landing page message match

Question: "Would someone who searched this term expect this page?" Answer type: Noul. Next: Mismatches feed the Quality Score fix list.

This is the semantic half of Quality Score. Status codes and load time stay with a crawler, the expectation check goes to Jev.

6. Recommendation triage

Question: "Apply, dismiss or review this Google recommendation?" Answer type: Choice. Next: Dismissals are applied, anything touching budget waits for a human.

Google's recommendations tab mixes useful fixes with budget raises. A Choice with the account's goals in the state sorts them, and the threshold keeps money decisions human.

7. Brand vs non-brand split

Question: "Is this query a brand search?" Answer type: Noul. Next: Reporting splits cleanly without a regex that misses misspellings.

Brand regexes break on typos and product names. A yes/no per query is slower to describe and far more accurate on the tail.

8. Disapproval routing

Question: "Is this disapproval a policy issue, a destination issue or an editorial issue?" Answer type: Choice. Next: Each one is routed to the right fix: copy, page or appeal.

Small job, daily value. The route is decided in milliseconds and the fix is written by the LLM or done by a person.

The pattern behind every use case

Every use case on this page follows one architecture. Code finds the candidates, Jev judges them, a threshold decides what happens, and a frontier model writes only what passed.

This is the same split Ryze AI already uses between changes it makes on its own and changes that wait for approval. A cheap decision model widens the first group without touching the rule for the second.

What it costs at the list price

These are arithmetic at TypeSafe's published price of $0.042 per 1M input tokens with free output. They are not measured runs, and token counts depend on how much context you put in each state.

The point of the table is the order of magnitude. Jobs that cost tens of dollars per run on a frontier model, and therefore run monthly, cost cents and can run nightly.

What Jev cannot do

Jev is narrow by design, and the narrow part matters as much as the price.

• No rationale. You get 0.91, never a sentence explaining it. Anything a client or a finance team will read still needs a frontier model to write the reason.
• Bounded answers only. If the set of possible answers is not known up front, it is the wrong tool. Strategy, copy and diagnosis stay with an LLM.
• Accuracy is triage-grade. 67.8% agreement on the vendor's own benchmark is fine behind a confidence threshold and wrong for auto-applying a budget change.
• "Cannot hallucinate" means it cannot break the schema. It can still pick the wrong option. Measure it against human labels on your own data before trusting a threshold.
• Waitlist and pricing risk. There is no general availability date, and the price may move.

Keep the frontier model for the search-term *rules* (what counts as a buyer for this business), for writing RSAs, for diagnosing a CPA jump, and for the sentence that explains each queued change. Keep a human on budgets and bids over 20%, new campaigns and conversion actions.

Where to go deeper on each job

A decision model only sorts. The guides below cover the work around it: getting the data, writing the rules, and making the changes.

• Search terms report
• Negative keyword lists
• Negative keyword mistakes
• Low ad strength on RSAs
• Conversion tracking
• Budget pacing rules
• Claude MCP automation for Google Ads
• AI keyword research for Google Ads
• Analyzing search terms with AI

How to start before you have access

Jev is in early access. There is no public self-serve signup yet.

The Ryze help center documents what the Ryze agent manages in Google Ads and approvals.