The five failure modes we see in audit after audit, the order they have to be solved in, and the thirty-day repair plan a competent in-house team can run before they call a firm. If you have run the queries that matter for your category and your firm is absent, the cause is almost certainly on this list.
If a B2B firm is missing from ChatGPT, Perplexity, Gemini or Claude answers, the cause is almost always one of five failure modes, in this order of frequency: (1) entity fragmentation · the firm is described differently across the open web; (2) unliftable evidence · claims live as marketing prose, not as named, dated, structured artefacts; (3) citation desert · no presence in trusted publications, registries or professional bodies; (4) answer geometry mismatch · the page does not begin with a direct answer to the buyer's question; (5) retrieval blockers · robots.txt, JavaScript-only rendering or paywalls keep the model out of the page entirely. The thirty-day repair runs in this exact order.
Diagnose first · the three queries to run before anything else
Before you fix anything, prove the symptom. Most firms believe they are absent from AI answers because a colleague reported a single search came back without their name. A single query is not a diagnostic; it is an anecdote. Run three queries instead, in three different surfaces, and write down what you see.
- The category query. "Who are the leading [category] firms in [geography] for [buyer profile]?" This is the question your buyer asks. Look at whether the answer names you, in what position, and how the model characterises you.
- The branded query. "What is [your firm]?" This is the entity question. Look at whether the description is correct, complete, and consistent with what your homepage says today.
- The comparison query. "Compare [your firm] to [competitor]." This is the depth question. The model can only run this if it has enough information about both firms. If it cannot, that is itself the diagnosis.
Run all three on at least three surfaces · ChatGPT, Perplexity and Gemini. Save the answers. The patterns of failure across them point at which of the five modes below you are stuck inside.
Failure 1 · Entity fragmentation
The single most common reason a B2B firm is invisible inside generative models is that the model has read three or four different versions of who they are and has chosen, defensibly, not to commit to any of them. The website calls the firm one thing. LinkedIn calls it another. Crunchbase has a third entry. The press kit uses a fourth name. The category label drifts.
The model's job is to identify a single canonical entity it can reason about. Confronted with fragmentation, the model degrades · it either refuses to commit, or it commits to whichever fragment has the highest signal density inside the substrate it is currently reading.
How to identify it
- The branded query returns a vague or contradictory description of the firm.
- The model says some version of "there are several firms by that name" or "I do not have detailed information about this firm."
- The category label the model uses is not the one the firm uses internally.
Why it persists
Entity fragmentation usually accumulates over time. A firm rebrands but the old name lives on in old press releases. The category label evolves but the registrar entry never updated. Founders publish under personal handles that do not link back to the firm. None of these are individually wrong; together they are corrosive.
Failure 2 · Unliftable evidence
The second most common failure is that the firm has the experience but has not exposed it in a form the model can lift. The work is real. The case studies exist. The methodology is mature. None of it is in a structure the model can quote.
Liftable evidence has three properties · named, dated, and structurally exposed. A firm that says "we have helped many clients" has nothing the model can cite. A firm that says "we ran the 4E System with twelve enterprise SaaS partners between 2022 and 2025, with a documented 47% lift in qualified pipeline within nine months" has a paragraph the model can lift verbatim into an answer.
How to identify it
- The category query lists competitors with named methods (the "X System," the "Y Framework"); your firm is described in generic terms.
- The model cannot quote any specific outcome, partner type, geography, or methodology when asked about your firm.
- Your case studies live as PDFs, locked behind a form, or as images of slide decks rather than as structured HTML pages.
The named-method principle
Naming a method is not a marketing flourish. It gives the model a noun to attach to the firm. An unnamed methodology is mechanically invisible: there is nothing for the model to remember and quote. A firm that runs an unnamed but excellent practice will lose, every time, to a firm with a named, mediocre one.
Failure 3 · Citation desert
The third failure is structural to the open web, not to the firm. The firm has no presence in the corpora the model treats as authoritative. There are no industry-publication citations. No professional-association inclusion. No standards-body listing. No academic mention. The firm exists only on its own domain and a handful of low-trust syndications.
This is fixable, but it is the layer that takes longest. A model's trust in a citation source is built up over years of corpus-wide patterns; you cannot shortcut it with a paid placement. BrightEdge has tracked the impact of Google's AI Overviews on click-through behaviour across their enterprise client base · the directional finding is that AI-mediated answer surfaces compress traffic for publishers that are not cited in the answer, and lift visibility for those that are. The premium for being inside the answer, rather than below it, is the practical case for citation density.
The trust hierarchy
| Tier | Source type | Relative weight |
|---|---|---|
| Tier 1 | Standards bodies, professional associations, academic publications, regulator-adjacent journals | Highest |
| Tier 2 | Major industry analyst outlets, credentialled trade press, recognised conferences | High |
| Tier 3 | Established business publications, founder-led publications, named podcasts with archives | Medium |
| Tier 4 | Press release wires, paid syndication networks, content-marketing platforms | Low |
| Tier 5 | Low-tier directory listings, scraped aggregators, anonymous review sites | Minimal · sometimes negative |
A single Tier 1 citation often outweighs hundreds of Tier 4 mentions. Marketing teams instinctively start at Tier 4 because it is purchasable and measurable. The model rewards Tier 1 because it is unpurchasable and adversarially vetted.
Failure 4 · Answer geometry mismatch
The fourth failure is internal to the firm's own pages. The model retrieves a small reading window · typically a few thousand tokens · when it grounds an answer. What that window contains determines what gets cited. If your top page opens with a hero, then a video, then a "story" section, then a carousel, and only reaches the substantive content at the bottom, the model's reading window has already closed by the time it gets there.
Answer geometry is the discipline of putting the answer where the model can find it. One question per page. Direct answer near the top. Evidence beneath. Cross-links to adjacent answers.
The four properties of good answer geometry
- Single-question page. The page exists to answer one buyer or model question, not to summarise the whole firm.
- Canonical answer in the first 300 words. A direct, structured paragraph the model can lift verbatim · not a sales pitch.
- Named evidence beneath. Methods, partners, dates, outcomes. Structured data where applicable.
- Internal links to adjacent answers. Each page is a node in a graph; the graph signals depth.
Failure 5 · Retrieval blockers
The fifth, and most embarrassing, failure is that the firm has accidentally locked the model out of the building. This is the rarest failure but also the easiest to verify, so check it first. Look for any of these:
- A
robots.txtthat disallows AI crawlers ·GPTBot,Google-Extended,CCBot,PerplexityBot,ClaudeBot,Applebot-Extended· by default. Many marketing teams added these blocks during the first wave of "AI scraping" panic in 2023–24 and never reversed them. - A site that renders entirely client-side, with no server-side HTML for the model to read. Some models cannot execute JavaScript at retrieval time.
- Paywalled or login-gated content, including case studies and methodology pages locked behind a "download" form.
- A site that returns 200 to humans but a soft-block, anti-bot challenge or aggressive rate limit to known model crawlers.
- Pages that exist but are not in the sitemap and have no internal links pointing at them · orphan pages the model cannot find.
If any of these are true, no amount of work on the other four layers will help. The model has to be able to read the page before it can cite the page.
The thirty-day repair plan
The order of operations matters. Skip a row and the rows below will not produce results because the substrate is still broken.
| Day | Task | Why it has to come first |
|---|---|---|
| Day 1 | Diagnostic queries | Run the three queries · category, branded, comparison · across ChatGPT, Perplexity and Gemini. Save the outputs. Without a baseline you cannot measure progress. |
| Day 2 | Retrieval audit | Check robots.txt, server-side rendering, paywalls, sitemap. Unblock anything blocking AI crawlers. This is mechanical and can be done in an afternoon. |
| Days 3–5 | Entity reconciliation | Pull every public reference. Fix the inconsistencies. Pick a canonical name, category, founding date, leadership, headquarters. Update the registrar, LinkedIn, Crunchbase, partner directories. |
| Days 6–8 | Schema layer | Add Organization, ProfessionalService, LocalBusiness (where relevant) schema to the homepage. Add Article and Person schema to leadership pages and any blog content. |
| Days 9–12 | Name your method | If you have a methodology, name it. Publish a single page that defines it · its components, its sequence, its applicable scope. Date it. |
| Days 13–17 | Publish one liftable case | Take the strongest engagement you can publish. Write it as a structured case note: question, framework, outcome. Named partner where allowed. Dated. With Article schema. |
| Days 18–22 | Restructure the top 5 pages | Each of the five most-visited pages: one question per page, direct answer in the first 300 words, named evidence beneath, internal links to adjacent answers. No redesign · restructure. |
| Days 23–25 | Citation outreach round 1 | Identify ten Tier 1 / Tier 2 publications in your category. Pitch a substantive, well-evidenced contribution to two of them. This is the longest-lead investment in the plan; start it now so it lands on Day 60+. |
| Days 26–28 | Re-run diagnostic | Run the same three queries. Compare to Day 1 baseline. The first wins on entity and retrieval should be visible by now; citation wins will not yet be. |
| Days 29–30 | Set up the measurement loop | Schedule a monthly probe of the same queries. Track citation share and position quality. This becomes the dashboard the firm runs against from Day 31 onwards. |
How to know it is working
Forget impressions. Forget traffic. The two metrics that matter for AI visibility are citation share and position quality.
- Citation share. Of the ten queries that matter for your category, in what fraction is your firm named in the answer? Track it monthly across ChatGPT, Perplexity, Gemini, Claude.
- Position quality. When you are cited, are you the first name in the answer, the second, or the fifth? Is the characterisation positive (the standard answer for X) or generic (one of several options)?
A useful baseline shape: at the start of a thirty-day repair, citation share for most B2B firms in our partner portfolio is between 0/10 and 2/10. By Day 30, the entity and retrieval fixes alone typically push that to 3/10 to 5/10 with mid-to-low position quality. The citation-density investment, started on Day 23, lands between Day 60 and Day 120 and pushes the position quality up to first or second name. Beyond Day 120, the firm is in a different conversation.
What comes after the thirty days
Day 31 is not the end of the work. It is the start of the operating discipline. The model's index is moving · pre-training cycles, retrieval-corpus refreshes, system-prompt updates. Your position is not a static asset; it is a portfolio that has to be checked.
What good looks like, ninety days in:
- Citation share above 6/10 across the queries that matter for the category, on at least two surfaces.
- Position quality at first or second name on the highest-intent queries.
- A monthly cadence of checking, logging, iterating · owned by a single person inside the firm or its agency, not buried inside a "marketing dashboard" nobody reads.
- One Tier 1 or Tier 2 placement from the citation outreach round that started in Days 23–25.
If the thirty-day repair is run honestly and the firm still does not move, the cause is almost always external · a category that is genuinely contested, a competitor with a far longer head start, or a model that has formed an outdated impression of the firm that requires more aggressive corpus correction. That is the conversation we have inside a Strategic Growth Audit, and the diagnostic playbook above is also released · fork-and-use · in Open Source.
FAQ
How do we know which of the five failure modes we are stuck inside?
Run the three diagnostic queries · category, branded, comparison · across ChatGPT, Perplexity and Gemini. The pattern of failure across surfaces points at the cause. A vague branded answer is entity fragmentation. A confident answer that lists competitors but not you is citation desert. An answer that mentions you generically without specifics is unliftable evidence. A model that says it cannot find a current page is a retrieval blocker. A model that finds the page but cites a competitor's homepage instead is answer geometry.
Can the thirty-day repair be done by an in-house team or do we need an agency?
Most of it can be done in-house by a competent marketing or content team with access to the website and the registry credentials. The exceptions are the citation outreach (which requires editorial relationships in the relevant publications) and any deeper schema or rendering changes (which may require a developer). The first three weeks are entirely doable in-house.
Will fixing AI visibility help or hurt our Google rankings?
Help. The fixes that make a page liftable into AI answers · clean schema, valid HTML, fast first paint, direct-answer geometry, internal linking · are the same fixes that make a page rank well in traditional search and surface in Google's AI Overview. The two disciplines have a roughly 80% overlap; the work compounds across both surfaces.
What does "citation share" actually look like as a measurement?
It is a fraction. The numerator is the number of category queries on which your firm is named in the answer; the denominator is the total number of queries you are tracking. A firm tracking ten queries that is named in three is at 3/10. The metric is tracked per surface (ChatGPT, Perplexity, Gemini, Claude) because the surfaces have different retrieval indexes and different system prompts, and the fix that lifts citation share on one may not lift it on another.
Is unblocking AI crawlers in robots.txt safe? Won't it expose our content?
Yes, it is safe · and necessary if you want to be cited. Allowing AI crawlers does not give them more access than a regular search engine has; they read the same public HTML. What changes is whether the model can ground its answers in your current content. Blocking AI crawlers does not protect content; it ensures the model answers without you in the picture.
Get an honest read of your site.
Drop your URL and a work email. Within five working days we send back a read built on your domain, your category and the buyers asking about you today · including which of the five failure modes is currently costing you visibility, and the order to fix them. No follow-up sequence, no newsletter.