Testing AI search tactics: what works and what doesn’t

Ten common AI search and GEO tactics, each tested on a brand-new site with verified crawler logs. The claim, the test, the result, and a verdict: supported, not supported or inconclusive.

On this page
  1. How each test works
  2. The scoreboard
  3. Early observations
  4. What we expect
  5. What we will not test
  6. Limits
  7. About this article
  8. Get the next Census by email

Short answer: Most AI search advice is untested. We are testing ten common tactics on a brand-new domain, one variable at a time, with verified crawler logs and dated citation checks. Each tactic gets a verdict: supported, not supported or inconclusive. Results so far are below, and the table is updated as tests finish.

How each test works

Every tactic is tested the same way: the claim as it is usually made, the setup that tests it, the result in numbers, and a verdict. “Inconclusive” is a normal outcome.

The scoreboard

#TacticThe claimHow we test itStatus
1llms.txtAI crawlers read it and use it to find your best pagesLog every request for llms.txt and check who made itRunning since 29 Oct 2026
2Markdown versions of pagesAI tools prefer clean Markdown to HTMLLog every .md request and every request that asks for MarkdownRunning since 29 Oct 2026
3IndexNowGets pages into Bing, and so into Copilot, fasterTwo articles published at the same moment, one submitted to IndexNow and one not. Compare first crawl and first citationRunning since 5 Nov 2026, 09:00 UTC. First crawl so far: the IndexNow article, 27 seconds after its submission (by YandexBot, which cannot be verified); this article, none yet
4JavaScript-rendered contentAI crawlers do not run JavaScript, so they miss itTwo visible test markers on one page: one in the HTML, one added by JavaScript. See which ones engines quotePlanned
5FAQ structured dataHelps AI tools lift your answersMatched articles with and without FAQPage dataPlanned
6Content SignalsAI crawlers change behavior when you set ai-train or ai-input to noFour weeks with all signals on, then one switched off, comparing visits by crawler classRunning since 29 Oct 2026
7Google top 10You must rank in Google’s top 10 to be cited in AI OverviewsCompare our AI Overview citations with our Search Console positionsRunning since 29 Oct 2026
8BacklinksA page with no backlinks will not be citedOur first article has no backlinks. Track whether and when it is citedRunning since 29 Oct 2026
9Question headingsHeadings phrased as questions get quoted word for wordCheck AI answers that cite us for verbatim heading textRunning since 29 Oct 2026
10LinkedIn postsSharing on LinkedIn speeds up the first Bing and Copilot crawlShare one of two matched articles on LinkedIn, then compare first crawl timesPlanned

Last updated: 5 November 2026, 09:25 UTC.

Early observations

On day 0, the first crawler to request our llms.txt file and the Markdown versions of our pages was AhrefsBot, an SEO tool’s crawler, not an AI company’s. Verified crawlers followed later that day: ClaudeBot fetched a Markdown copy at 15:37 UTC and llms.txt at 15:38 UTC, and Google’s GoogleOther crawler fetched both that evening. The full three-day count is in Do AI crawlers read llms.txt? That is not a verdict on tactics 1 or 2; the tests run for at least 30 days.

What we expect

Written before any results: llms.txt and Markdown versions will see little or no verified AI crawler traffic in the first month (tactics 1 and 2). IndexNow will speed up Bing crawling measurably (tactic 3). We have no prediction for tactic 4, because AI crawlers differ on JavaScript and their behavior has changed over time.

What we will not test

Anything that only works by deceiving a search engine or an AI model: showing bots different content from people, hidden text, instructions planted for AI models, fake freshness, bought links. Some of these may “work” for a while. They are not worth the risk to any real site.

Limits

One site, a handful of pages, and engines that change from week to week. A “not supported” verdict here means the tactic made no measurable difference on this site in this window. It does not mean it can never work. Every experiment on this site is listed on our experiments page.

About this article

This article is half of tactic 3. It was published at the same moment as How long does it take a new website to get cited by AI search? and deliberately not submitted to IndexNow; that article was. Its test reference is Census test marker LN-004.

Published by the AI Visibility Census.

Newsletter

Get the next Census by email

Each new edition, list and experiment result, sent when it publishes. About one email a week. One click to leave.

We send a confirmation email first. If you add a website, we may tell you when your company appears in a Census result. Privacy notice