Contents5
A good share of SEO work is not writing. It is a string of closed decisions, repeated thousands of times: should this page link to that one? Is this query informational or commercial? Does this page deserve an update, a merge, or nothing at all? Is this site worth a link?
That is exactly the ground covered by AI decision models, a category opened in September 2026 by Jev, from TypeSafe AI. A widely shared LinkedIn post lists seven SEO uses, with numbers. The core idea is right; several figures and one promise deserve a closer look. Here are the seven uses, what to take from them, and how we put them in place.
In short
A decision model does not write: it picks from a list, scores against a scale or judges true or false, with a probability.
At Jev’s price (0.042 dollars per million tokens read, no output charge), 10,000 decisions of about 1,000 tokens each cost around 42 cents. The figure holds.
The test results going around come from the vendor or its users, not from an independent measurement: rerun them on your own data.
The confidence score tells you what a human should review, not that you can skip the review. Any decision that deletes, merges or sends something gets approved.
Why SEO suits decision models
A large language model writes word by word: it is slow, and you pay for the output. A decision model reads the input in one pass and returns a typed answer your code can use directly. Jev’s documentation offers three question types: a choice from a list, a score against a scale, and a yes or no with its probability. Each answer comes with a confidence value.
SEO is full of questions in that shape, asked at volumes nobody reviews by hand: hundreds of pages, thousands of queries, dozens of sites to assess.
The seven SEO decisions to hand over
- Internal linking.
For each page, the model picks the best target from a shortlist of pages. The test quoted claims 586 pages read in 45 seconds and 584 links placed for 21 cents. The principle is the one in our 10-minute internal linking method, at scale: the model suggests, the anchor rules and the review stay yours.
- Sorting keyword exports.
Each query is classified: informational, comparison, purchase or off-topic. You only review the low-confidence cases. It is the safest of the seven, because a mistake only costs one misfiled line.
- Content audits.
Each page is matched with its last three months of Search Console data, then sorted: keep, update, merge or delete. Deleting or merging stays an option, provided it is approved by the site’s strategy and checked by a human: a page with few visits can still carry links, a conversion or a topic you want to cover.
- Filtering sites for backlinks.
Each site is scored: off-topic, relevant or highly relevant. It saves real time. On the other hand, “checking 700 messages in 40 seconds” pushes towards mass sending: an outreach email that looks like 699 others ends up as spam, whichever model checked it.
- Pre-publication review.
Intent, sources, internal links, title against H1: every question goes in a single request, and a failed check sends the draft back for review. It is an excellent safeguard, provided the questions are precise.
- Choosing the model for each task.
A small update goes to a fast, cheap model; a pillar page to the most capable one. This is not specific to Jev: it is model routing, and it is often where the AI bill is decided.
- Mapping your answers for AI engines.
50 real customer questions crossed with 100 pages give 5,000 scored cells: complete answer, partial or missing. The unanswered questions become your list of pages to write to get cited by answer engines.
What the advertised numbers are worth
The cost can be checked. Jev’s price list shows 0.042 dollars per million tokens read, with no output charge. A decision that reads one page or one export line rarely exceeds 1,000 tokens, hence 42 cents for 10,000 decisions. Other decision models are cheaper still, as we detail in our comparison.
The results, however, come from public demonstrations, not from an independent measurement. “584 links placed” does not say how many were good. Before running a job across a whole site, we draw a sample of around fifty decisions, check it by hand, and set the confidence threshold from the error rate we observe.
Watch out
The model does not decide for you
The post says Jev “makes the decisions for you”. True for sorting a list of queries, false for anything that commits the site. Deleting a page, merging two pieces of content, sending a message: the model prepares the decision, the strategy approves it, a human checks it.
What makes the result good
The model is only one layer. Quality comes from three things it does not supply:
- Good input data. An up-to-date list of pages, real Search Console data, clean keyword exports. A decision model rules on what it is given, nothing more.
- Well-written questions. “Does this page answer question X?” gives a clear result; “Is this page good?” gives none.
- A rule before the model. What a simple rule can settle (a page returning 404, an empty title) needs no AI. The model handles the grey areas, and a human handles what falls below the confidence threshold.
How we put it in place
A decision model is only one tool among others: we explain where it fits in what an AI-assisted SEO agency should deliver. We start by listing the SEO decisions that recur for you, and their volume. Then we write the questions, test them on a hand-checked sample and set the thresholds. Finally, we connect everything to your data (Search Console, page list, exports), with a “to review” queue for uncertain cases.
That is the idea behind our AI at the right cost offer: the right model for each task, a rule when a rule is enough, and a cost that follows your real volume rather than a subscription per tool. And when search visibility itself is the goal, our SEO and GEO offer builds these sorting jobs into the site’s ongoing work.
Sources
- Jev’s documentation (new tab)docs.typesafe.ai
- Jev’s price list (new tab)docs.typesafe.ai
Jim · founder of Brain Plus AI
Also in seo and ai visibility
Understand29 min
AI decision models: Jev, Clef, uses and costs
Jev, Clef and their rivals write nothing: they decide, with a probability attached. What these decision models are used for in each line of work, how to put them in place, what they really cost and what we make of them.
Practical guide4 min
Internal linking: find your links in 10 minutes, no tool needed
Internal linking is the cheapest SEO lever, and the one most often done badly. The eight-step method to find the links to add with a simple Google search, the two rules worth nuancing, and how we automate it once a site passes a few dozen pages.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.