Method

The Gane Index records what AI models say when a buyer asks which finance product to use. It is not a review site, hands-on testing or a collection of customer ratings. A first-choice share is not market share, revenue or customer satisfaction, and it does not show that a product suits a particular buyer.

Results are dated observations. Products, cited pages and model behavior change, so check the facts that matter before buying, and do not rely on the Index alone.

This area covers 57 categories across seven functions, each asked for three company sizes. A standing always belongs to one category and one company size, and nothing is pooled across sizes. A product missing from a category has not been judged negatively: no model may have named it, it may be under the eligibility rule below, or a name may have been misread.

Each monthly edition asks fourteen models from twelve labs, each lab's low-cost model, and two from Meta and OpenAI. The edition page lists them. A new model version is treated as a new model and measured again.

Every category is asked six ways for each company size, which makes 84 answers per category and size. Each question gets a fresh session, so nothing carries over from an earlier answer. Models are asked through their APIs with web search available, not through consumer assistant apps, so an answer in an app may differ. Search being available does not mean it was used; in this edition 89% of answers used it. An answer that fails after retries is left out rather than filled in; shares count the answers collected.

These are examples, not the full prompt set; each category page shows its own questions. The negative question asks for warnings, so a warning from it was requested rather than volunteered. Pages name the question beside each quote.

A judge model reads each answer with its question and category. For every product named it records the name as written, a label and a quote of the evidence. A product counts only if the answer treats it as a candidate in the category asked about. Category names, analyst firms, people and names appearing only inside a link are not counted.

From the November 2026 Edition the judge is a model trained for this task, on the earlier judge's labels and on answers that human readers settled. Before the change it was compared with the earlier judge on 13,073 answers it had not been trained on: the same leader in all 52 categories, with the largest share gap per product at a median of 1.1 points and at most 3.1. Each edition, 5% of answers are read again by Claude Opus 5 under the same rules. Earlier editions keep the labels they were published with.

Automated reading can misidentify a product or misread its sentiment. A quote shows what a model wrote. It does not establish that the statement is true, and a citation does not establish that the cited page is right.

Product names are resolved through a versioned vendor table of names, aliases and parent companies. A bare company name counts as its product in the asked category only when it has exactly one there. That is an assumption, since a model may mean the company generally, so each such reading is listed on the category page. Parts of a combined answer count separately when each part resolves. Names that are not products in the category, such as general-purpose tools and publishers of software rankings, are excluded by rule. The category list gives what each category covers. A change to the table can change how earlier answers read, as described under Editions.

Models often answer the same question differently when it is asked twice. So each edition repeats a fixed sample of categories within a week, with the same questions, models and settings. In the October 2026 Edition repeat, 68% of first choices changed. Live search results and the models' own variation both contribute, and the repeat does not separate them.

The floor is the 90th percentile of how far a category leader's share moved between the two runs: 10 points in the October 2026 Edition. A smaller change is reported as within the floor. That does not prove nothing changed; it means the change can't be told apart from repeat variation. A new leader is reported only if it clears the old one by more than the floor. Movement between editions is read over the models both editions asked.

The models read the web, and the Index is on the web. In the November 2026 Edition, 8.5% of answers cited an Index page. Within each model, those answers chose October 2026's leader by 30.1 points more often than other answers. Applied across all answers, that is worth about 2.6 points of a leader's share, against a floor of 10 points. This does not show that a page caused the difference. The Index reports the gap every edition.

The Gane Index is published by Gane, which sells marketing software, assessments and execution services and may compete with products represented in the Index. That puts the publisher in the market it measures. No product of Gane's falls in a category this area covers; the disclosure is made anyway, here and in the footer.

One method applies to every company. No company can pay to change a label, a share, a rank or a verdict, and being a Gane customer changes nothing here. Gane's customers, trials and prospects are kept out of the Index's pipeline, and the build fails if a private run reaches a public page. Gane is in the vendor table: a model naming it is published like any other label, and so is its absence.

A vendor can claim a listing and correct what its own site states. Those statements are labeled as the vendor's and never change a measured result. Inclusion, or the use of a name or logo, does not indicate endorsement by that company. The labs and suppliers that are also measured here are listed on the conflicts page.

Editions publish monthly and each is kept at its own address (every edition). Answers and judge labels are stored as collected. When the vendor table changes, earlier editions are read again under the new table, so an archived page can differ from what it first showed; the change log names each version and each material correction.

To challenge a finding, write to index@gane.ai with the page and the figure. No claimed listing is needed. Credible reports of serious factual errors, misidentification and other significant harm are prioritized. A material correction identifies what changed, when and why, and does not overwrite the original record.

Cited sources show what a model retrieved, not what it was trained on. The categories are the Index's own judgment and vendors may dispute their placement. Whether a lab favors its own products can only be measured where that lab has products in the categories. Comparisons across a change of judge or model are read against the floor.