Methodology

This page explains how every Exibran Index report is made, so you can judge the numbers and repeat them. The rules and definitions stay the same from one Index to the next.

What changes each time (the dates, the questions, the AI apps) is listed for each edition at the bottom.

Who Runs It, and Who Pays?

I, Michael Zacs, run the research through Exibran. Nobody sponsors it, and nobody pays for any result.

Exibran sells AI search visibility to B2B SaaS companies, so I have a business interest in this topic. I say so on every report.

If I write about a company I have a link to, I say that in the report too.

What Rules Does Every Index Follow?

• I publish the questions I ask and the dates I ask them. The questions are in the Questions tab of each edition's data file.
• I publish the raw data of every Index as a download.
• I never accept payment from a brand I rank, and I never take money to change a result.
• I do not say one thing caused another unless I have tested it. A snapshot of what AI says is not proof of why.
• I report ranges, not single numbers, and I call overlapping ranges a tie.
• I correct errors in public, with a dated note in the changelog below.

What Do I Measure?

Three things, always kept apart:

• Mention: the tool appears anywhere in the answer, even in a list.
• Pick: the answer's own verdict chooses the product. A list that stops without a verdict is not a pick.
• AI Pick Rate: picks ÷ answers, shown with a 95% range (Wilson interval).

The full definition is in the AI Pick Rate guide.

Which Questions Do I Ask?

Buyer questions, written in the words a buyer would use, not search keywords.

I wrote them while setting up Radar. I first researched the buyers of each category: who the ideal customer is, who the competitors are, and where those buyers ask for help, including communities, forums, groups, and search.

Then I wrote questions a buyer would ask, matched to the stages they go through, with a focus on high buying intent. In every Index they fall into three groups:

• Neutral questions. They name no tool. These produce every headline rate in the Index.
• Questions that name tools (for example, comparing two). I report them separately as the "name effect."
• Questions that name a tool I do not rank. They are left out of every rate.

How Are Answers Collected?

Every Index is collected with Exibran Radar, my own software. Each question is asked three times on each AI app, on one date, from one country and in one language.

The edition notes below name the apps, the date, and the collection route for each app.

If an ask fails or an app shows no answer, I say how I handled it in that edition's notes.

How Are Answers Read?

Radar's AI reader tags each answer first: which tools it mentions, in the order it mentions them, and which one it picks. The order is saved in the data file.

The November 2026 report uses it to show which tool each answer names first. Then I run a full review pass over the results with AI help and fix the tags.

I do not have a second person tag the same answers, so I have no agreement score between two reviewers. Treat the tags as one reviewer's judgement.

After the review, I re-ask a sample of the questions in the real apps and compare them with what Radar saved.

That shows how much answers move between asks. It does not test whether I tagged the saved text correctly.

Site types in the sources section are assigned by the site name, so small or new products I could not match fall into "other".

Radar saves the site, not the page link. The "How AI Describes Each Tool" counts use simple word rules on the sentences about each tool, so treat them as approximate. They show what AI says, not what is true.

What Else Do I Check?

For each report, I also search the questions on Google in a logged-out private window (incognito) and record the top 3 normal results and their type.

How Do I Report Differences?

• Every rate comes with a 95% range. Where the ranges overlap, I call it a tie.
• When I compare one Index with an earlier one in the same category, I report movement under 7 points as no change.
• Percentages are rounded to whole numbers in text. The data file has the exact counts.

What Are the Limits?

• Three asks per question per app. AI answers change from ask to ask, so small gaps are noise.
• One date, one country and one language per edition. Results can differ elsewhere and on other days.
• Some apps are collected through an API or a scraper, which may not match exactly what a person sees in the app.
• One reviewer, with AI help.
• The questions are mine, and I built them from my own research into each market. A different list could give a different ranking.
• An Index is a snapshot of what AI says, not a measure of which product is best.

How Do I Correct Mistakes?

If a tool is listed wrongly, a number is off, or you are a vendor and think an answer was tagged wrong, email chat@exibran.com.

I will check the saved answer, fix the report if I got it wrong, and note the change in the changelog below.

What Did Each Edition Cover?

November 2026: AI Search Visibility Tools (Run C-011)

• Report: Index: Which AI Search Visibility Tool Does AI Recommend? November 2026. Data file: download.
• Collected: 3 November 2026, from the United States (country set to US), in English.
• Questions: 30 buyer questions about AI search visibility tools, also called AI search optimization tools. 18 are neutral, 9 name tools and 3 name Aiso. The 3 Aiso questions are left out of every rate.
• Answers: 30 questions × 3 runs × 5 AI apps = 450 answers. The headline rates use the 270 neutral answers.
• Routes:
  • ChatGPT: Bright Data, 90 answers.
  • Gemini: Bright Data, 90 answers.
  • Google AI Mode: Bright Data, 90 answers.
  • Google AI Overviews: Apify (apify~google-ai-overviews-scraper), 87 answers.
  • Claude: Anthropic API with web search on (model claude-sonnet-5-5), 90 answers.
• Failed asks: Google AI Overviews showed no answer for 3 of the 90 asks. I counted these as answers where no tool was mentioned. Leaving them out moves the neutral pick rate from 15.6% to 15.7%. Eleven early Claude asks failed because of a low credit balance, and I asked them again.
• Review: in my review pass, the tool list changed in 399 of 450 answers and the pick changed in 112. Most list changes were cleaning up names, such as merging Otterly.AI and OtterlyAI, and removing AI apps and bots that are not tools.
• Sample re-ask: on 4 November 2026, I asked 10 of the questions again in the real apps, once each. The tool lists matched or mostly matched in 3 of 10, and the pick agreed in 3 of 10. That is expected, because AI answers change every time.
• Google check: all 30 questions, on 4 November 2026.
• Not included: Perplexity and Microsoft Copilot are not in this edition.
• Conflict of interest: this edition names Aiso, a SaaS product I do not work with. I have no relationship with them yet, and nobody asked or paid for the report. I included it to show how a smaller product appears when AI is asked about it by name and when it is not.

Changelog

• 1.0, 4 November 2026: first version. General rules and definitions, plus the notes for the November 2026 edition.