GEO Tasks to Stop: Retiring Unsupported Items From Your Roadmap

Most GEO roadmaps carry tasks that sounded sensible two years ago and never had evidence behind them. The evidence now splits cleanly. The levers that hold up are access levers: server-side rendering, per-bot robots.txt permission, network reachability, index eligibility and snippet controls. They work by removing blocks.

A critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) concluded that already-retrieved content can change how it gets cited, but no reviewed technique shows a stable, cross-platform causal effect on discoverability. Almost every published GEO effect size came from a lab setup where the page was already in the model's context. That is the generation stage. Most tasks on your roadmap claim to affect the earlier stages: whether a crawler can fetch the page (access) and whether it enters the answer at all (retrieval).

That mismatch is the whole problem. A task tested at one stage and sold for another is a bet, not a plan. If you want the background on why studies point in different directions, read why GEO studies disagree. The next move is to sort every roadmap item by the stage it claims to change.

These are the items that show up most often in GEO roadmaps and agency proposals. The verdicts follow primary sources, not vendor blogs.

One boundary is worth keeping in view. Every page in the Ahrefs schema study already had 100+ AI Overview citations. Whether schema helps a page enter the answer set at all is untested, not refuted. That still means nobody can promise it works. For a longer list of claims to question, see GEO claims that don't hold up.

Block two hours with whoever owns SEO and whoever owns the site's code. Then work through these steps.

1. Export every open item whose rationale mentions AI, GEO, AEO, LLMs or AI Overviews. Include agency deliverables and recurring tasks, not only new projects.
2. Write the claim in one sentence. For example: "Adding FAQ schema will get us cited in AI Overviews." If nobody can state the claim, the task has no reason to exist.
3. Tag the stage it claims to affect: parametric (what the model knows without the web), access, retrieval, or generation. Be strict. Most content tactics only have evidence at generation.
4. Find a primary source. That means the platform's own documentation or a controlled study. A blog post repeating an old rule does not count, even if it is recent.
5. Assign a verdict: keep, reframe, retire, or park. Park means the question is open and nothing urgent depends on it.
6. Move the freed hours to access checks (see the section below). Do not let them drift back into another unsupported tactic.
7. Log each decision with the source and the date you read it. Platform documentation changes, and you will need to recheck it next quarter.

Retiring a task means you stop expecting AI visibility from it. It does not mean you undo it.

• Existing schema stays. Google says existing markup does no harm. Ripping it out costs engineering time and can break rich results that still exist. The honest position: it is not a lever for AI answers.
• An llms.txt can stay. Google's guidance is scoped to Google Search, and it says maintaining one for other services that use it is completely fine. Just stop reporting it as a visibility win.
• Real author information stays. Named authors and credentials can be sound content practice. What you drop is the idea of an E-E-A-T score and any KPI built on it.
• Freshness is unresolved, not dead. One 252,000-trial factorial study (arXiv 2605.25517) found explicit prices and a recent timestamp helped consistently in its test setup. Updating real information is fine. Bumping a date with no change is gaming your own measurement.
• Clear facts on fetchable pages stay. Stating a specific price, number or one-line definition on a page crawlers can read gives the engine a sentence it can quote. That is the only positive experimental signal for on-page content.

If you want a fuller picture of what does have support, read whether any proven way exists to increase AI citations.

The tasks worth the freed hours are mechanical and checkable. They do not guarantee citations. They remove reasons you would be excluded.

• Render key content on the server. Vercel found that ChatGPT and Claude crawlers fetched JavaScript files (11.50% and 23.84% of requests) but did not execute them. Pricing or product facts that only appear after JavaScript runs may be invisible to those crawlers. Googlebot is the exception, because it renders.
• Set robots.txt per bot. OpenAI states that each of its crawler settings is independent. Blocking GPTBot stops training use. OAI-SearchBot controls ChatGPT search. A blanket AI block can remove you from ChatGPT search without anyone noticing.
• Test the firewall and CDN. A WAF can return 403 to a crawler that robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
• Find accidental suppression. Look for stray noindex, nosnippet, max-snippet:0 and data-nosnippet inherited from templates.
• Confirm index status engine by engine in Google Search Console and Bing Webmaster Tools.

The code-level issues that stop AI indexing are the place to start, and a site migration can reintroduce them overnight.

Run this list against every AI-related item before the next planning cycle.

• The task states a specific claim in one sentence.
• The claim names the stage it affects: parametric, access, retrieval or generation.
• A primary source supports the claim at that stage, not only at generation.
• The source was read this quarter, with the date logged.
• The task does not rely on schema, FAQ schema, llms.txt, an E-E-A-T score, domain authority, chunking or AI-friendly writing as a reason for AI citations.
• The task does not report an AI rank or position in Google.
• Robots.txt rules are set per bot, with training and search bots handled separately.
• Key facts appear in the raw HTML before JavaScript runs.
• The CDN or firewall lets each allowed AI crawler through.
• No unintended noindex, nosnippet or data-nosnippet sits on important pages.
• Retired items are marked "no longer an AI lever", not deleted from the site.
• Freed hours are assigned to a named access task with an owner.

Good results

• The roadmap is shorter, and every surviving AI item names a mechanism and a primary source.
• Access work sits at the top, with owners and dates.
• Existing schema and author pages are untouched.
• Reports stop using phrases like "AI rank" or "E-E-A-T score" and say what was checked and what was fixed.
• Stakeholders hear plainly what you control (access, accuracy of your own pages) and what you do not (what the models choose to say).

Bad results

• Old myths got swapped for new ones with fresher branding.
• Someone removed working markup to prove a point.
• A retired metric came back under a new name.
• The team promised a citation count or score in exchange for the reallocated budget.
• Access checks were added to the list but never run.

If the outcome looks like the second list, go back to step 4 and ask for the primary source again. The AI SEO best practices sorted by evidence is a useful reference for that conversation.

Run the full review once a quarter. Platform documentation moves fast, and a task you retired could gain evidence, or a task you kept could lose it.

Also run it when any of these happen:

• A platform changes its crawler or AI features documentation.
• Your CDN changes defaults. Cloudflare set Training and Agent crawlers to blocked by default on ad-displaying pages for some customers from 15 September 2026, without site owners editing anything.
• You migrate the site, change frameworks or change hosting.
• A vendor or agency pitches a new tactic with a percentage attached.
• A new controlled study lands that tests getting into answers, not only behaviour inside them.

Fold the quarterly review into a wider plan, such as a 30/60/90 day AI visibility plan, so pruning and fixing happen on the same calendar.

Should we delete our existing schema markup? No. Google says existing structured data does no harm, and some rich results still depend on it. What changes is the reason you keep it. Schema is not a lever for AI answers: Ahrefs' matched test found no major uplift and a 4.6% drop in AI Overviews citations after adding it. Keep the markup, stop reporting it as AI work, and stop adding new markup on the promise of citations.

Is publishing an llms.txt harmful? There is no evidence it harms you. Google says Search ignores the file, so it neither helps nor hurts there, and it calls maintaining one for other services completely fine. No other engine documents reading it either. If you already have one, leave it. Just remove it from your AI visibility roadmap and from any report that credits it with results.

Do author bios and credentials still matter? They can be good content practice, because readers want to know who wrote a page. What does not exist is an E-E-A-T score. Google states E-E-A-T isn't a specific ranking factor, so there is no number to improve or track. Keep honest author pages, and drop any KPI, dashboard or vendor report that claims to measure E-E-A-T.

Should we stop updating old content? No. Freshness is an open question, not a retired one. One large factorial study found explicit prices and recent timestamps helped in its test setup, but the main correlational study had no control group. Update pages when the facts change, such as pricing or features. Do not change dates without changing content, because that only distorts your own measurement.

How do we push back on an agency that sells these tasks? Ask for the one-sentence claim behind each deliverable and the primary source that supports it at the stage it claims to affect. Platform documentation and controlled studies count. Vendor blogs repeating old rules do not. If the answer is a percentage uplift from a lab study, ask whether the pages were already inside the model's context. Most were.