Not "consider improving". A ranked list, with the evidence for each item.

Prescriptive Fixes for AI Search

AI search optimization starts with removing what blocks you. After every audit, 99Visibility turns what it found into a ranked fix list: what to change, why, how urgent it is and roughly how long it takes.

Much of the advice sold as AI search optimization is additive: add markup, rewrite pages, publish more. The best-designed test of citation-oriented rewriting found most methods "largely ineffective" and often harmful (C-SEO Bench, arXiv:2506.11097, NeurIPS 2025).

What does have a documented mechanism is access: whether a crawler may fetch your page, whether your firewall lets it through, and whether a snippet control hides your text. Those fixes are ranked by the consequence the vendor documents, so a ChatGPT search block is critical and a training-only block is low.

The list also covers wrong facts, the sources engines cite instead of you, and the prompts where competitors are named and you are not.

Each fix carries its evidence: the answer it came from, the line in your robots.txt, or the vendor document that says what a block does. You should be able to defend every item to a developer or a client without taking our word for it.

What the fix list covers

Crawler access

robots.txt is read for 18 AI crawlers, and your page is requested as each crawler that fetches pages, to see what your firewall does. A block that removes you from AI search ranks high or critical; a training-only block ranks low and says it does not affect AI search.

Snippet controls

noindex, nosnippet, max-snippet:0 and data-nosnippet keep text out of AI answers. When one is on your page, for example inherited from a template, the fix says where.

Wrong facts

Accuracy flags become fixes that quote the sentence the error came from and tell you where to state the correct fact.

Facts that used to be true

When an answer repeats something that changed, like a product you sold, the fix tells you to state the current fact once, dated, and to ask the cited sites to update. It also says what a page cannot do: a model answering from memory keeps the old fact until its next version.

Missing mentions

For prompts where engines skip you, or name a tracked competitor instead, the fix lists the sites they cited, so you know which pages shape the answer.

Markup that disagrees with the page

If your structured data states a fact your visible text does not, the fix tells you to make them agree. It never tells you to add markup to get cited.

How it works

  1. 1

    The audit collects the evidence

    The answers, their citations, your accuracy flags, competitor mentions and the access check of your site all come from the same run.

  2. 2

    Rules turn it into fixes

    Each fix has a title, the action, why it matters, a priority from critical to low, and an effort of about 30 minutes, 1-2 hours or 3+ hours.

  3. 3

    The list is ranked

    Every rule reads the same audit, and fixes with the same title are merged into one. Each category is capped, with more room for critical and high items, so one problem cannot flood the list; what is left is sorted from critical to low.

  4. 4

    You work it

    Every plan gets the fix list. Starter ($49/mo) and up adds a step-by-step guide under each fix.

  5. 5

    You confirm it

    Mark a fix done when it ships, then re-run the audit or the access check to see whether anything changed.

A real fix, from the rule that writes it

When robots.txt blocks OpenAI's search crawler, the access check produces this fix.

Example output

Title
ChatGPT search cannot see your site — OAI-SearchBot is blocked in robots.txt
Priority
Critical
Effort
About 30 minutes
Action
Add a group to robots.txt that re-opens the site for search: "User-agent: OAI-SearchBot" followed by "Allow: /". The fix names the line in your file that causes the block.
Why
OpenAI, verbatim: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links." (https://developers.openai.com/api/docs/bots). This is independent of GPTBot: you can keep training blocked and stay eligible for search.

The title, priority, effort and quotation are what the rule produces; the action is shortened here. The offending line is filled in from your own robots.txt.

From a real audit: Under Armour’s fix list

On 15 October 2026 we audited Under Armour with Nike, Adidas and Puma as tracked competitors. These are fixes the rules wrote from that audit, in the order the engine ranked them, with their titles as written (one has been reworded since; see the note).

From a real audit

PriorityEffortFix
Critical3+ hoursUnder Armour is absent for "best running shoes 2026"
Critical3+ hoursZero visibility on "best running shoes 2026", your gap with the highest estimated demand
High1-2 hoursKnown from training data, missed by live retrieval for 1 prompt
MediumAbout 30 minutesAI repeats outdated information about Under Armour
Medium1-2 hoursYour prompts cluster around these topics — check you have a page for each
MediumAbout 30 minutesOld facts about Under Armour are still being repeated

The two critical items are one gap seen twice: no platform named Under Armour for "best running shoes 2026". The first lists the sites the answers drew on instead, such as runnersworld.com, runrepeat.com and rtings.com, and says to get Under Armour onto those pages.

The high item is a split between engines: Claude, answering from memory, named Under Armour for "best soccer cleats", and Perplexity, which searched, did not. Its fix is about access, whether the page can be fetched and read, not about how it is written.

The two outdated-fact items come from the MyFitnessPal error, raised when the saved answers were re-scored with the fixed accuracy check.

Eleven fixes came out, and Google AI Overviews was not part of the run. Left out here are three whose titles rest on the Gemini column, its answers cut short by a token limit on our side, and two competitor items produced by a known bug. The second item’s title is the one the rule writes today: on the day it described the prompt by its traffic, but the number behind it is our estimate from the prompt’s wording and measures nothing, so the fix now calls it estimated demand.

Read the AI visibility audit of Nike, Adidas, Puma and Under Armour

Fix it with Agent V

Agent V is the assistant beside your dashboard, and this is where it does most of its work. Pick a recommendation and ask V to fix it: it drafts the page copy, the robots.txt lines or the JSON-LD to paste, depending on what the fix needs.

When the fix is done, V can mark the recommendation done for you, as a card you confirm.

Anything that changes your account, like running an audit or adding a keyword, comes to you as a card to confirm. Nothing changes until you do, and V never edits your site. It runs on Anthropic's Claude models and is on every plan.

How Agent V works

Who it is for

  • Marketers without an SEO specialist who need a list they can hand to a developer.
  • Developers who want the exact line to change and the vendor document that says why.
  • Agencies that have to explain each recommendation to a client and defend it.

What it does not do

  • It does not tell you to add markup or rewrite pages "for AI" to get cited. Google's own guide says "there's no special schema.org markup you need to add" (AI optimization guide, updated 2026-07-10).
  • It does not promise a mention. Removing a block makes a page eligible to be fetched; whether an engine then uses it is the engine's call.
  • It cannot prove a firewall block from outside. We cannot send requests from the vendors’ own IP addresses, so a pass is not proof, and a refusal can come from an impostor check that would let the real crawler through.
  • It does not make the change for you. Fixes are instructions that you or your developer ship.
  • Some fixes are checks, not findings. We do not render your pages, so "confirm the facts are there with JavaScript off" is a step for you to run.

Questions

How is a fix prioritised?

By its documented consequence, not by how alarming it sounds. A ChatGPT search block is critical because OpenAI documents the result; a training-crawler block is low because it does not affect AI search.

What is SEO for AI search?

For Google, it is SEO: its guide says "optimizing for generative AI search is optimizing for the search experience, and thus still SEO". Its AI features page adds: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." That is why this list starts with whether a page can be crawled, indexed and quoted.

How do I get ChatGPT to recommend my business?

Nothing can make it, but a block can make it impossible. OpenAI says "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers", and recommends allowing the bot in robots.txt and allowing requests from its published IP ranges. The access check tests both: what robots.txt tells OAI-SearchBot, and what your firewall returns when we request the page as that crawler.

Do the fixes work?

A fix removes a documented reason you cannot appear; it does not promise that you will. Re-run the audit after you ship it: answers that search the web can change once the engine fetches the page again, and answers given from memory change only with a newer model.

What does the Free plan get?

The full fix list for each audit and the AI crawler access check. Starter ($49/mo) adds the step-by-step guide under each fix and the detail of each accuracy flag.

Why do some fixes send me to other sites?

Engines that search the web build answers from the pages they retrieve. The fix lists the domains they cited for your prompts, so you know where being listed matters.

See What AI Sees

Start on the Free plan. No card needed.