The LLM visibility chain is the six conditions a brand must meet in sequence to be recommended in an AI answer: accessibility, entity recognition, category match, proof of condition, external confirmation and accurate framing. The chain is only as strong as its weakest link. If one link is missing, strength in the others doesn’t carry into the answer, so visibility work starts by finding the earliest broken link.
Picture two brands. The first has a site full of rich content, but its security settings block AI crawlers. The second has an accessible site and good content, but its company name is spelled three different ways in external sources. Their problems differ; the result is the same: neither is in the answer.
Visibility doesn’t depend on a single attribute; it depends on a series of conditions all being met. Recro calls this series the LLM visibility chain. The chain model breaks “why aren’t we showing up?” from one big problem into six questions that can be checked in order.
AI knows a brand through two routes
To understand the links, first separate the two routes through which a model reaches information.
Training data. Large language models learn from text collected up to their training cutoff. How a brand sits in the model’s “own knowledge” depends on how much, and how, it was mentioned on the web in the past. This route changes slowly.
Live search. ChatGPT’s search feature, Perplexity, and Google’s AI Overviews and AI Mode search the web at answer time. Google has said AI Mode uses a “query fan-out” technique that splits a single question into many sub-queries. This route changes quickly and is decisive for most B2B decision questions.
All six links apply to both routes, but with different weights. Accessibility and proof of condition matter most in live search; entity recognition and external confirmation are critical in both.
The six links
- Accessibility. Can the crawlers AI tools use reach the pages? robots.txt rules, firewall and CDN settings, JavaScript-only content or login-gated pages can break this link.
- Entity recognition. Does the model recognize the brand as a single, consistent entity? Different spellings of the company name, other companies with the same name or old trade names weaken this link.
- Category match. Does the model place the brand in the right category? Brands that describe themselves with generic phrases like “solution partner” or “technology company” struggle to match specific category questions.
- Proof of condition. Is information answering the question’s condition (certification, capacity, integration, region, pricing model) clearly published? This link is where Content and Communication GAPs happen.
- External confirmation. Do the brand’s claims appear consistently in independent sources? This link is where the Authority GAP happens.
- Accurate framing. When the model recommends the brand, does it describe it with the right reasons and current information? An old product, the wrong service area or wrong pricing can turn visibility into damage.
Accessibility: which crawlers, which settings?
This link is the most technical, the most often skipped and the most costly when broken: if the crawler can’t reach the page, none of the following links work.
| Crawler | Owner | Used for |
|---|---|---|
| GPTBot | OpenAI | Collecting content for model training |
| OAI-SearchBot | OpenAI | Surfacing sites in ChatGPT search results |
| ChatGPT-User | OpenAI | Fetching a page on a user’s request |
| PerplexityBot | Perplexity | Citing sources in Perplexity answers |
| Googlebot | Search, AI Overviews and AI Mode | |
| Google-Extended | A token controlling use of content for Gemini model training |
OpenAI’s documentation says these crawlers can be managed separately in robots.txt: a site can allow OAI-SearchBot and disallow GPTBot. So a preference of “don’t use my content for model training” doesn’t have to mean “don’t show me in ChatGPT search.” But the distinction has to be made deliberately.
There’s been a shift on the infrastructure side too. Since July 1, 2025, Cloudflare blocks AI crawlers by default on new domains; site owners must explicitly allow them. While marketing produces content, it may not know that a hosting or CDN setting keeps these crawlers out.
Entity recognition: who does the model think the brand is?
A model recognizes a brand to the extent it can merge information from different sources under a single “entity.” Google’s Knowledge Graph, open knowledge bases like Wikidata and consistent directory records make that merge easier. The sameAs field in a site’s Organization markup is one technical step that links the brand’s LinkedIn page, directory records and other official profiles to the same entity.
Brands in Turkey carry an extra risk: suffixes like “Şirketi”, “A.Ş.” or “San. ve Tic.” and Turkish characters written differently across sources can lead the model to treat one company as two entities. We covered this in exporting brands.
Why is the chain only as strong as its weakest link?
The links work in sequence. An inaccessible page never reaches the model, however good its content. If the model can’t recognize the brand, it can’t place it in the right category. Without a category match, proof of condition is never examined.
That’s why investing in a link at the end of the chain is ineffective while an earlier link is broken. A brand trying to build external confirmation while its site is closed to crawlers spends much of its budget on a broken chain.
How do you find the broken link?
| Link | Symptom | First check |
|---|---|---|
| Accessibility | The brand is never cited as a source in any searching tool, for any question | robots.txt, firewall and CDN settings, readability without JavaScript |
| Entity recognition | The model confuses the brand with another company or says it has no information | Whether the name is spelled the same on the site, in directories and on LinkedIn; Organization and sameAs |
| Category match | Recognized when asked by name, missing when asked by category | Category definition on the homepage and service pages |
| Proof of condition | Present in discovery questions, missing in conditional ones | Whether the questions’ conditions are clearly written in page text |
| External confirmation | The model describes the brand cautiously or vaguely | Number and consistency of records in independent sources |
| Accurate framing | The brand is in the answer but with wrong or old information | Old pages, distributor content, outdated directory records |
Checks should run in order from start to finish. Once the first broken link is found, later links can only be read accurately after it’s repaired.
Why is accurate framing the last but most expensive link?
The last link is when the brand enters the answer but is described wrongly. A product discontinued three years ago is still recommended, the brand is credited with a region it doesn’t serve, or old pricing is passed on. When the buyer comes to sales with that information, the first meeting starts with corrections. In healthcare and regulated sectors, misrepresentation carries a bigger risk than invisibility. This link can only be measured by reading the content of the answer; a simple “was the brand mentioned?” count can’t see it.
How does Recro use the chain?
The Recro Insight Model reads simulation results through the chain model. For every question where the brand is missing, the broken link is identified and the action plan is ordered from the start of the chain to the end, so investment goes to the earliest broken link first.
One simulation · 4 key pillars · 90+ steps
What sets Recro apart is four pillars built from scratch for each brand. Together they form a single simulation of 90+ steps, with each pillar producing the input for the next.
- Persona builder: Decision-maker profiles that represent the brand’s real buyers, specific to its sector and sales structure.
- Question builder: The brand-specific decision questions these people ask AI along their buying journey.
- Report builder: A report that reads the answers against the brand’s goals and competitors, together with the sources behind them.
- Action recommendation builder: A prioritized list of actions that closes the sector-specific signal gaps.
How the pillars contribute in the chain model:
- The question builder places question types that test different links into a natural buying flow: entity questions asked by brand name, discovery questions asked by category, conditional screening questions and verification questions that probe risk. The question type where the break begins points to the broken link.
- The report builder reads answers as content, not just presence or absence: the links that source the conversation, the order in which the brand enters the answer and the accuracy of what’s said. The accurate framing link can only be measured through this reading.
- The action recommendation builder prioritizes findings in chain order. If there’s a technical access problem, the list starts there; content and external source actions follow.
We describe the measurement method in how to measure AI visibility.
Sample finding format · illustrative
- Symptom: Recognized when asked by name; absent from category and condition questions.
- Check: Site accessible; entity consistent; homepage describes the company as an “integrated solution partner”.
- Broken link: Category match
- Action: Rewrite the category definition on the homepage and service pages in the terms buyers use; proof-of-condition work starts after that.
Executive summary
- Entering an AI answer depends on six links being met in sequence. The chain is only as strong as its weakest link.
- Accessibility is the most often skipped link: crawler settings and CDN defaults can keep content from ever reaching the model.
- Checks run start to finish and investment goes to the earliest broken link. The last link, accurate framing, can only be measured by reading the content of answers.
Frequently asked questions
Do we have to allow AI crawlers?
No, it’s a choice. But blocking the crawlers used for search and answer generation largely prevents appearing as a source in those tools. Training crawlers and search crawlers can usually be managed separately.
How do we know whether our site is open to AI crawlers?
By checking robots.txt rules, firewall and CDN settings, and these crawlers’ requests in server logs. If responses return 403, the crawler is being blocked.
What strengthens entity recognition?
Spelling the company name the same in every source, updating directory records, Organization markup with a sameAs field linking official profiles, and keeping the LinkedIn company page consistent with the site.
Which link is fastest to repair?
Accessibility and proof of condition can be repaired relatively quickly with on-site changes. External confirmation and entity recognition in training data take time.
Is the chain the same in every sector?
The links are the same, but their weights differ. Accurate framing matters more in regulated sectors, entity recognition for exporters, and external confirmation in professional services.
Which link in your chain is broken? Recro’s brand-specific simulation finds the broken link in every question where you’re missing and prioritizes actions in chain order.
Request a demo insight report →
Sources
- OpenAI · Overview of OpenAI crawlers
- Cloudflare · Permission-based approach for AI crawlers (2025)
- Search Engine Journal · Query fan-out technique in AI Mode
- Google Search Central · AI features and your website
- Schema.org · sameAs
This article was prepared with AI assistance and published under the review of Mehmet Semih İpek.
