Tag
Start here
Fine-tuning is the one capability that makes an open decision model genuinely different from a hosted one. Here is how to build a label set from your own review decisions, train on a laptop, and avoid the trap that inflates every published benchmark.
Tag
A Jev internal-link run proposed 679 links across 566 pages in under six seconds for 27 cents. An editorial pass kept 287 of them. Here is the full attrition, the rubric rewrite that improved recall from 45% to 65%, and how to run the same benchmark on your own site.
The vendor benchmark says 67.8%. An independent reranking test says Jev is not a free win. Here is how to read both, why the numbers are not comparable, and how to run your own accuracy test before you trust a threshold.
We ran 20 prompts across three AI surfaces and tracked where the sources came from. Five prompts found 31 percent of the source universe, ten found 54, and the twentieth still added 14 new domains. There was no plateau.
Google publishes no endpoint that returns your rank. We measured what its actual APIs do return: data that is three days old, an average position that differs from a live SERP result by 7.4 positions at the median, and an inspection verdict that says NEUTRAL for a URL Google has never seen.
We sent one prompt to the same model six times with web search on. Across the six answers, 18 distinct URLs were cited and not one of them appeared in all six runs. Here is what an AI visibility tracker has to record if it is going to mean anything.
An AI Overview tracker only earns its place if it measures presence, position, and citations over time. We ran the first snapshot across 40 SEO queries. Ten percent had an AI Overview, and the pattern was not the one the panic suggests.
Free SERP checkers stop at position 100. We asked Google for 200 results across six queries and it returned 83 to 128 organic results, then stopped on its own. Here is what actually sits past position 100, and when it is worth checking.
A claim sweeping search teams says putting one brand first in a comparison query makes AI Overviews recommend it up to 80% of the time. We tested that with 102 queries across 16 brand pairs. Order changed nothing. Here is what did.
A study of one million pages finds that AI involvement does not prevent indexing or high rankings. The practical question for content teams is how they use AI time to improve evidence, judgment, and page quality.
