The AEO Tactics I Will Not Recommend

In May 2026, 97% of llms.txt files on the web received zero requests. Not one fetch, by any bot, for a full month. Ahrefs measured it across server logs from 137,210 domains.

People spent real hours generating those files. Someone sold a few of them as a deliverable.

That is the pattern worth naming. Call it the checklist tax. Work that is easy to sell, easy to complete, easy to report, and measurable at zero. Five items on the standard AEO checklist have been tested and failed. Here is the evidence, so you can stop paying for them.

Does llms.txt do anything?

No. Two independent studies say so.

SE Ranking analyzed about 300,000 domains in November 2025 and found llms.txt on 10.13% of them. They ran Spearman correlation and an XGBoost model against citation frequency. No measurable relationship. Removing llms.txt made their model more accurate.

Ahrefs then went to the logs. Of the llms.txt files they found live, 97% got no traffic at all in May 2026. Among the few that did see requests, the largest single requester group was SEO audit tools, not AI engines. GPTBot accounted for 4.51% of those fetches.

Google’s John Mueller said it plainly in June 2025: no AI system currently uses llms.txt. OpenAI’s own crawler documentation names robots.txt and never mentions llms.txt.

Adoption kept climbing anyway. In my crawl of 3,877 B2B software sites, 1,041 of 3,233 publish one. Thirteen of the 34 sites whose firewalls blocked ChatGPT outright also publish an llms.txt, which is a welcome mat on a locked door.

If you already published one, leave it. The file costs nothing to host. Stop maintaining it, and stop paying anyone to update it.

FAQ and HowTo schema do not buy citations

No. The one controlled test available says the opposite.

Ahrefs added schema markup to 1,885 pages and tracked them against roughly 4,000 control pages for 30 days in May 2026. AI Overview citations fell 4.6%, a small but statistically significant decline. AI Mode moved up 2.4% and ChatGPT 2.2%, both indistinguishable from zero.

Plenty of teams shipped it anyway. In my crawl, 247 of 3,236 homepages carry FAQPage markup and 423 carry SoftwareApplication or Product markup.

Schema still earns its place for one job. Organization markup with accurate sameAs links tells a machine which company you are, which matters when three companies share your name. That is entity disambiguation, and it is worth doing once.

Adding schema types to chase citations is not a strategy. It is a deliverable wearing a strategy’s clothes.

Google-Extended is not a crawler

No. Google-Extended is not a crawler.

It is an opt-out token for training use. It will never appear in your server logs, because nothing fetches with it. AI Overviews retrieve from the live Googlebot index, so blocking Google-Extended changes nothing about whether you appear in one.

The same confusion runs the other direction with OpenAI. GPTBot collects content for training. OAI-SearchBot fetches your page to answer a question and cite you. Block GPTBot and you have made a business decision about training. Block OAI-SearchBot and you have cut off the live fetch, so ChatGPT answers from whatever older copy sits in its index rather than from your current page.

Do more G2 reviews raise your citation rate?

Barely. A profile matters enormously. Volume past a profile does almost nothing.

Kevin Indig analyzed 500 G2 categories and 30,000 AI citations for G2 itself, published October 2025. Review volume came out statistically significant and practically tiny: an R-squared of 0.009 for citations, 0.012 for share of voice. His own summary is that reviews explain under 2% of the variance.

Then look at presence. Seer Interactive studied 804,491 AI responses across 1,926 brands in March 2026. Brands with no verified review profile were cited 1% of the time. Brands with as few as 1 to 13 reviews were cited 53.5% of the time.

Review platforms do get cited. SE Ranking found 34.5% of AI Overviews on commercial queries cite at least one, with G2 holding 23.1% of those links and Capterra 17.8%.

So the instruction is narrow. Get a complete, current profile on the platforms your category uses. Categorize it correctly. Then stop. A review generation campaign aimed at AI visibility is spending budget on the flat part of the curve.

Bing indexation is a gate, not a lever

Being in Bing’s index matters. Your rank inside it is not the thing to manage.

Seer Interactive examined about 500 SearchGPT citations and found 87% matched a result in Bing’s top organic positions. So Bing indexation behaves like an eligibility gate, the same way a review profile does. Not indexed, rarely cited.

The gate is also cheap to clear. Verify the site in Bing Webmaster Tools, confirm the pages you care about are indexed, and stop. A campaign to move Bing rankings is not a campaign to earn ChatGPT citations, and tracking Bing positions as a visibility KPI will tell you almost nothing about whether you are being named.

Four fixes that serve both channels

Four fixes serve Google ranking and AI retrieval at the same time. They are the ones to approve first.

  1. Original media. Your screenshots, diagrams, charts and data, in place of stock images.
  2. A named author with a real bio, a resolvable author page, and Person markup.
  3. Methodology shown rather than asserted, with dates and sample sizes on the page.
  4. Answers in the first sentence under each question-shaped heading, so a passage survives being lifted out.

Most of the rest of the AEO checklist does not have that property. Some of it trades Google traffic for AI visibility. Naming the four fixes that cost nothing in organic is the cheapest way to get a content budget approved by someone who only measures Google.

Before any of it, confirm two things. Confirm retrieval crawlers can reach your pages. Confirm your primary claim sits in the initial HTML. The fix order matters more than the fix list.

Ask this before you approve the scope

Put one question to any agency or consultant pitching AEO work. What will you refuse to recommend, and what evidence makes you refuse it?

A provider with no refusals is selling a checklist. The checklist is cheap to produce and expensive to work through, and you pay for both. What a real audit should include covers the rest of the questions worth asking.

The audit publishes its refusals with the studies behind them. It scores 30 fixed sections on whether machines can reach you, read you and select you. It runs 150 observations across ChatGPT, Google AI Mode and Perplexity, so you know where you stand.

See what the audit covers.