Your 30-Day AI Visibility Fix Plan

You now know more about how AI answers get made than most people selling GEO. The risk at this point is doing everything at once, or starting with the work that looks most like work: rewriting pages, adding markup, publishing an llms.txt. None of that has evidence behind it. Access fixes do.

This capstone turns the course into 30 days of work, ordered by strength of evidence. Week 1 measures. Week 2 opens access. Week 3 fixes the facts. Week 4 checks each engine and sets up monitoring so the gains do not quietly break. Every row gets an owner, a date and a line that says when it is done.

Why subtractive fixes go first

An access fix removes a reason you cannot be used. A crawler blocked by robots.txt, turned away by a firewall, handed a page whose facts only exist after JavaScript runs, or told nosnippet: each is a binary failure, and the engines' operators document the consequence themselves.

So the order is: fixes with a documented mechanism first, fixes with a plausible mechanism second (making your own facts consistent), and no time at all on drop-list tactics. Two more rules shape the month:

• Change one layer at a time. Access in week 2, facts in week 3. If everything ships on one day, you cannot tell what moved what.
• Do not expect training-data answers to change. An answer given with web search off comes from a model frozen at a knowledge cutoff, and that is how 99Visibility measures ChatGPT and Claude. Nothing you ship this month reaches those answers. Judge the month on the Perplexity, Gemini and AI Overviews columns and on your leading indicators.

Week 1: Measure

You cannot say a fix worked without a baseline taken before it.

• Run the baseline audit with a frozen prompt set of category, comparison and brand-fact prompts, three runs, logged out. See Run Your First AI Answer Audit, read the result per platform as Understanding Your Visibility Score explains, and record the variance band.
• Write the brand fact sheet: price per plan, features, features you do not have, founding date, headquarters, CEO and leadership, ownership, acquisitions and headcount. Enter it in your 99Visibility brand profile so answers are checked against it, as in Find What AI Gets Wrong About You.
• List every wrong answer by type, platform and prompt.
• Record the analytics baseline: the GA4 AI referral segment, Search Console AI impressions and, if you are verified there, Bing AI citations, as set out in Monitoring and Measuring GEO ROI.
• Build the competitor grid from Competitor Analysis for GEO.

Done when: the band, the fact sheet, the error list and the analytics baseline are written down and dated.

Week 2: Open access

• robots.txt, bot by bot. Allow the retrieval bots: OAI-SearchBot, PerplexityBot, Claude-SearchBot, Meta-WebIndexer, Googlebot and bingbot. Decide on training bots separately; blocking GPTBot does not remove you from ChatGPT search. Meta's is the only first-party statement that allowing a crawler helps citation: "Allowing Meta-WebIndexer in your robots.txt file helps us cite and link to your content in Meta AI's responses." Details in AI Crawlers and robots.txt, Bot by Bot.
• Firewall and CDN. Run the Crawler Access check and fix every 403, 429 or challenge served to a retrieval bot. If you are on Cloudflare, review its AI crawler settings, whose defaults Cloudflare announced it would change for some sites from 2026-09-15. Allow verified bots by user-agent plus a verified-bot program, not by IP lists alone. See When robots.txt Says Yes and Your Firewall Says No.
• Rendering. For every page that states a fact on your sheet, fetch the raw HTML without JavaScript and search for the fact. This check is manual, because 99Visibility does not diff raw against rendered HTML. See Rendering, noindex and Snippet Controls.
• Suppression. Remove any unintended noindex, nosnippet, max-snippet:0 or data-nosnippet. In Search Console, check the generative AI inclusion control, including any child property that inherits an opt-out from its parent.
• Bing. Verify the site in Bing Webmaster Tools and confirm your key pages are indexed.

Done when: every retrieval bot gets a 200 on your key pages, every fact on the sheet is in the raw HTML, and nothing is suppressed by accident.

Week 3: Fix the facts

• Sweep your own site. Search it for every fact on the sheet: old prices, retired plan names, features you dropped. Give each fact one canonical page that states it in plain, server-rendered text, and make every other page agree. This is the plausible mechanism behind "ChatGPT gets our pricing wrong". See How to Create Content AI Can Quote.
• Check schema parity. Every fact in your JSON-LD must match the visible text. Do not add schema to get cited. Use the schema generator's parity check and Schema Markup for AI: What It Still Does.
• Make claims checkable. Named authors and linked sources, as reader practice rather than a score, per E-E-A-T for AI: What Actually Counts.
• Request off-site corrections. Start with the third-party pages engines cited in your audit and your competitor grid. Send each a correction with the canonical URL, and log the date. See Building Off-Site Corroboration.

Done when: the site states each fact one way, schema matches the page, and every correction request is logged.

Week 4: Engine checks and monitoring

• ChatGPT: OAI-SearchBot allowed and reachable, key pages in Bing's index, tests run logged out. See Optimizing for ChatGPT.
• Perplexity: PerplexityBot allowed and reachable. Perplexity says its user-triggered fetcher generally ignores robots.txt, so do not count on robots.txt to control it. See Optimizing for Perplexity.
• Google AI Overviews: pages indexed and snippet-eligible, the Search Console control left on. There is no robots.txt token for AI Overviews; your Googlebot rules are the control. See Optimizing for Google AI Overviews.
• Claude and Gemini: Claude-SearchBot allowed. Google-Extended cuts grounding for Gemini Apps as well as training, so block it only on purpose. See Claude, Gemini and the Rest.

Then re-run the audit with the same prompt set and settings, and read it against the week-1 band. Set up the standing monitors: a robots.txt diff, crawler status checks, a raw-HTML fact check after every deploy, a monthly Bing and Search Console export, and the GA4 segment. Write the report in three parts: what you fixed, what moved outside the band, and what you observed in traffic.

Done when: the re-audit is read against the band, the monitors are running and the report is sent.

Stop doing

Take these off the plan. Each one is popular, and each one failed when someone checked.

Stop	Why
Publishing llms.txt for AI visibility	Google Search ignores it, and no other engine documents reading it
Adding FAQPage or other schema to get cited	FAQ rich results are deprecated, and Ahrefs' matched-control study found no citation uplift from schema
Chunking pages, or rewriting in an "AI-friendly" style with long-tail variants	Google says you don't need to, and tested rewrites were null to negative
Chasing an E-E-A-T score or Domain Authority	Google says E-E-A-T isn't a ranking factor, and no engine publishes an authority score
Pinging IndexNow for AI citations	The protocol never mentions AI, and Google does not consume it
Allowing GPTBot "to get cited", or blocking it to leave ChatGPT search	GPTBot governs training only; OAI-SearchBot governs search
Tracking a rank inside Google's AI answers	Google's AI report has no position, and third-party tools have no ranking data
Bumping page dates to look fresh	Freshness is exposed to engines, not shown to be weighted, and the date is self-declared
Quoting "40% more visibility", "ChatGPT is 92% of AI traffic" or crawl-to-referral ratios	In-context lab tests, link-tagging asymmetry and stale 2025 snapshots

Also off the plan: building anything on review-profile "citation rate" studies. The best-known one never defines the metric.

Your plan template

The columns are the template: week, task, lesson, owner, due date and done when. One owner per row, and the owner is a person, not a department. Here it is filled in for Ledgerly.

Week	Task	Lesson	Owner	Due	Done when
1	Baseline audit, three runs	First audit	Maya, marketing lead	Oct 21	Band recorded per platform
1	Fact sheet into brand profile	Hallucinations	Maya	Oct 22	Every fact type entered
1	GA4 segment and competitor grid	ROI, competitors	Ana, growth	Oct 23	Baseline note dated
2	Unblock retrieval bots in robots.txt	robots.txt	Sam, engineer	Oct 27	Rules verified live
2	Clear firewall challenges	Firewall	Sam	Oct 29	Every retrieval bot gets a 200
2	Server-render the pricing table	Rendering	Sam	Oct 30	Prices in raw HTML
3	Remove free-plan mentions site-wide	Quotable content	Maya	Nov 3	Site search finds none
3	Schema parity on pricing page	Schema	Sam	Nov 4	Offer prices match the page
3	Correction requests to two listicles	Corroboration	Ana	Nov 6	Requests sent and logged
4	Engine checks and re-audit	Engine playbooks	Maya	Nov 13	Read against the band
4	Monitors running, report sent	ROI	Ana	Nov 18	Report in the finance lead's inbox

Notice what is not on it: no rewrite sprint, no llms.txt, no new schema types. Ledgerly's month is spent removing reasons it cannot be read, then making its own site say one thing.

Common mistakes

• Starting with week 3 because it feels productive. Rewriting pages that crawlers cannot reach changes nothing an engine sees.
• Shipping every fix on the same day. You lose any chance of telling what moved what.
• Giving engineering rows to marketing. robots.txt, the firewall and rendering need someone who can deploy.
• Rows without a done-when line. A task with no finish line never finishes.
• Judging the month on one re-audit. Read it against the band, and expect training-data answers not to move at all.
• Treating day 30 as the end. The next deploy or CDN default can undo week 2. Month two is monitoring.