How a B2B Company Can Audit Its Presence in LLM-Generated Responses

Your buyers are using ChatGPT, Perplexity, Claude and Gemini to build vendor shortlists before they ever visit your website. Most B2B companies have no idea whether they appear in those conversations or not.

The independent benchmarks agree on the scale of it. The 2026 2X AI Visibility Index, which analysed 70 B2B companies, found that only 4.3% maintain a healthy discovery funnel. The other 95.7% appear mainly in queries where the buyer already knows the company name, which means they are absent from the earlier conversations that decide the shortlist. A larger study from ThriveStack, covering more than 6,000 brand websites across eight B2B sectors, found 87% scoring below 40 out of 100 on AI visibility.

And the gap is not explained by weak SEO. Walker Sands found that the median enterprise B2B company ranks for nearly 9,700 keywords and is still cited in just 3% of the AI Overviews it is relevant for.

The distance between ranking and being cited is the B2B marketing problem of 2026.

An LLM presence audit tells you where you actually stand: which engines mention you, how accurately, with what sentiment, against which competitors, and through which sources. This article covers how to run one properly, what the numbers mean, and where the audit stops being useful.

Most B2B teams are measuring the wrong thing

Search rankings and AI citations are not the same signal, and treating them as one is the most common mistake teams make when they try to assess their AI presence.

A high ranking tells you Google indexed your page and that your links and on-page signals pushed it up the results. A citation tells you something else: that a generative system trusted your content enough to use it inside an answer delivered straight to the buyer.

The distinction: SEO makes your pages eligible. Whether an AI system actually uses them as a source when it answers your buyer is a separate question with separate causes.

The two diverge sharply, which is why plenty of companies in the ThriveStack sample rank well and score badly. The structural problems that suppress citations, missing schema, keyword-dense copy with no answer architecture, blocked AI crawlers, are largely invisible to standard SEO tooling.

The inverted discovery funnel

2X describes most companies as running an inverted discovery funnel: present when buyers search for them by name, absent from the earlier conversations where buyers explore a category and assemble a shortlist.

That inversion matters commercially because AI-assisted discovery is front-loaded. A buyer who asks an engine for the best platforms for their use case, and does not see you, may never reach the stage of searching for you by name. The shortlist is built before your site is ever opened.

Which is why an audit starts with category prompts, not branded ones. Branded mention rate tells you whether people who already know you can find you. Category mention rate tells you whether people who do not know you can discover you. Only the second one builds pipeline.

The five steps

Each step builds on the one before it. Skipping any of them produces a number without context, which is worse than no number.

Step 1: build a prompt library across four tiers

The prompt library is the instrument. It has to cover real buyer intent, not just branded queries.

Thirty to fifty prompts is a workable set. Write them in buyer language, not internal product language. Most B2B teams have never built one at all, so finishing this step alone puts you ahead of the majority of your category.

Step 2: run every prompt across all four engines

The engines disagree far more than most teams assume, and running only ChatGPT then extrapolating produces a false picture. Writesonic analysed 161,286 prompts across ChatGPT, Gemini, Perplexity and Google AI Overviews and found that only 3.8% of sources were cited by all four for the same prompt. Between any two engines the overlap ran from 12% to 24%, and 72% to 73% of cited domains appeared on exactly one engine and nowhere else.

So run each prompt on all four:

• ChatGPT, which for most B2B categories is where the volume is.
• Perplexity, the most useful engine for this work because it surfaces source URLs natively.
• Claude, which retrieves from a different index again and skews toward professional buyers.
• Gemini, which matters most where your buyers live inside Google Workspace.

Run each prompt at least three times per engine, in a fresh session, and record the mention rate as a rate rather than a yes or no. Variation between runs is not measurement error. It is how these systems work: a study of five LLMs run ten times each under settings intended to be deterministic reported "accuracy variations up to 15% across naturally occurring runs" and concluded that "none of the LLMs consistently delivers repeatable accuracy across all tasks, much less identical output strings". A single run is one draw, not a measurement.

Step 3: score five things per prompt

1. Mention. Did the brand appear at all?
2. Position. Named first, listed among options, or absent?
3. Sentiment. Positive, neutral or negative framing?
4. Accuracy. Is the description of the product actually correct?
5. Source. Which URL or domain did the engine reference?

The logging is tedious and it produces the only thing that matters later: a baseline to measure against once work begins.

Step 4: map the citation source footprint

Build a list of every domain the engines referenced when they mentioned you. The map shows two things: where your authority currently lives, whether that is owned content, press, review platforms, Reddit threads or directories, and where the gaps are, meaning the high-authority domains in your category that cite competitors and not you.

Perplexity is the best engine for this step because it shows clickable citations by default.

Step 5: benchmark against two or three competitors

A visibility number with no competitive context does not mean anything. Run the identical prompt set for your closest competitors and compare mention frequency by tier, category rate against branded rate, sentiment and accuracy differences, and citation source overlap.

One brand usually owns a disproportionate share of category mentions. Knowing whether that is you tells you whether you are defending a position or closing a gap.

What the audit tells you, and what it does not

After all five steps you will know your category mention rate per engine, your share of voice against named competitors on the same prompts, which engines know you and which treat you as invisible, whether the descriptions are accurate or hallucinated, and which third-party domains drive your citations.

What it will not tell you is why. It will not tell you which content changes would move the rate, or whether your citation sources are stable. On that last point there is a real number: Wellows found that when the same query is re-asked two weeks later, only 34.7% of cited sources stay the same. Roughly two-thirds churn. A one-off audit is out of date almost immediately, which is an argument for tracking rather than for auditing harder.

On what counts as a good number, be careful with benchmarks. There is no established industry figure for a healthy category mention rate, and any vendor quoting one is quoting their own sample. The honest version is comparative: your rate only means something against your own previous rate and against the competitors you measured on the same prompts, on the same day, in the same sessions.

The gap between knowing the score and moving it is where most teams stall. They run the audit, build the spreadsheet, and then face a list of structural problems spanning content, technical and off-site work.

The five disciplines that move the number

These are not independent tactics. Each one reinforces the others.

Prompt research

Before any content or technical work, you need to know which prompts your buyers actually run. This is not keyword research: the queries are longer, more conversational, and usually carry use-case context that keyword tools never surface. The output is the prioritised prompt set that becomes both your baseline and your tracking set.

Technical optimisation

AI crawlers are not Googlebot. Plenty of sites that Google indexes completely are partly or wholly inaccessible to AI retrieval because of WAF rules, CDN configuration or client-side rendering. The technical pass covers crawler access for the retrieval agents specifically, schema, answer-first structure on key pages, and freshness signals.

ThriveStack's sector table puts professional services at an average of 28 out of 100, with 91% below 40, and names the primary cause as keyword-dense copy with no answer structure. That is a content-shaped problem that only a technical audit surfaces.

Content

Citable content is built differently. Each page needs a direct answer at the top, a clear heading hierarchy, and sections that can be lifted without the rest of the page for context. This is not a volume exercise. One well-structured page that answers a category question directly is worth more than ten posts that bury the answer in paragraph four.

Community presence

Engines treat community discussion as third-party corroboration, and in B2B categories Reddit threads turn up in citations alongside owned content and press. The work is identifying where your buyers actually discuss the category and contributing genuinely useful answers there. LLMReach does not use fake accounts, automated comments, vote manipulation, undisclosed promotion, or content designed to imitate genuine community discussion.

Citation tracking

Without tracking you cannot tell whether any of the above is working. Run the agreed prompt set across all four engines on a fixed cadence, log mention rate, position and sentiment, and compare against the baseline. Weekly spot checks on your highest-priority prompts, a monthly full run, and a quarterly pass that refreshes competitive benchmarks is a cadence that holds up.

Start with the baseline

Most B2B companies with active marketing programmes, established sites and real content teams are structurally invisible in AI answers. It is not usually a resource problem. It starts with not knowing the baseline, and without one you cannot attribute pipeline to AI discovery, justify the investment internally, or tell whether anything you do is working.

The LLMReach free AI visibility audit establishes that baseline. During a guided review meeting, LLMReach walks you through your priority buyer prompts, current AI visibility, competitor citations, source patterns, and the technical or content gaps that matter most. You leave the call knowing where the gap is, what is causing it, and which changes would matter first.

Your audit is reviewed live on the call. It is not emailed as a PDF.

Where LLMReach runs the full Citation Stack, the guarantee is specific. No single input is guaranteed on its own: no individual page, platform, prompt or AI response is promised. The Citation Stack is guaranteed as a whole: a 30% increase in total citations across your site, measured against your own baseline, within 90 days. If that is not reached, you choose between continued work at no charge and a full refund.

The audit is the start, not the answer. But it takes one call, and it produces a prioritised gap analysis you can act on whether you work with LLMReach or not.

Frequently asked questions

How can a B2B company audit its presence in LLM-generated responses?

Build a prompt library covering awareness, category, comparison and problem prompts, run each one several times across ChatGPT, Perplexity, Claude and Gemini, and log whether the brand is mentioned, in what position, with what sentiment, how accurately, and which source the engine cited. That record is the baseline everything later is measured against.

Which metrics matter in an LLM presence audit?

Mention presence, position, sentiment, accuracy and citation source. Together they show not just whether the brand appears but how it is framed and which sources are producing the answer. Category mention rate matters more than branded mention rate, because it measures discovery by buyers who do not already know you.

Is it enough to test one AI engine?

No. A study of 161,286 prompts found only 3.8% of sources were cited by all four major engines for the same prompt, with 72% to 73% of cited domains appearing on exactly one engine. Testing only ChatGPT and extrapolating to the others produces a picture that does not hold.

What is the difference between an LLM audit and an SEO audit?

An SEO audit tells you whether a page can rank. An LLM audit tells you whether AI systems actually cite it in answers. They diverge: research on more than 6,000 B2B sites found 87% scoring below 40 out of 100 on AI visibility, and many of them rank perfectly well in traditional search.

How often should an LLM presence audit be repeated?

More often than most teams expect, because cited sources are volatile. When the same query is re-asked two weeks later, only about a third of cited sources stay the same. A one-off audit dates quickly, so treat it as the start of a tracking cadence rather than a deliverable.