LLM visibility: get cited by AI assistants

Why a single prompt tells you nothing

The most common way this work goes wrong is someone asking ChatGPT who the best agency in their city is, screenshotting the answer, and treating it as a measurement. It is not one. Ask the same assistant the same question twice in a row and the sources cited frequently differ. Ask two different assistants and the overlap is often close to nothing.

This is not a flaw to be engineered around. It follows from how these systems work. Retrieval varies between runs, the underlying search layers differ between products, and generation is not deterministic. A brand that appears in one run and not the next has not gained or lost anything in between.

So the unit of measurement cannot be a prompt. It has to be a set of prompts, run repeatedly, with results aggregated. Anything else is noise presented as a finding, and it leads to panic about a disappearance that was never real and celebration of a mention that will not recur.

• Identical consecutive queries often return different sources
• Citation overlap between different assistants is typically low
• Variation is inherent to retrieval and generation, not a signal of change
• One screenshot is an anecdote, not a measurement

Share of voice across a fixed prompt set

The measurement we run is share of voice. We agree a fixed set of prompts that reflect how your buyers would genuinely ask, usually between thirty and eighty depending on the size of the category. They cover recommendation questions, comparison questions, problem-led questions and category definition questions, because brands surface differently across those types.

That set is then run across the assistants your buyers use, on a fixed schedule, with each prompt run several times because of the variance described above. We log every brand named and every domain cited, not just yours. The output is a percentage: across all runs of all prompts, how often were you named, and how does that compare to the competitors in the same set.

The prompt set stays fixed once agreed, which is the discipline that makes the numbers mean anything. Changing the prompts changes the number without anything real having changed. When we do add prompts, we report the old set and the new set separately until there is enough history for the new one to stand on its own.

• Thirty to eighty prompts covering recommendation, comparison, problem and definition intents
• Multiple runs per prompt to average out the variance
• Every brand and domain logged, so competitor share is visible
• The set held constant, so month to month comparison is valid
• Results split by assistant, because they behave differently

What actually influences whether you get named

Assistants draw on two things: what was in training data, and what live retrieval returns when they search. You can influence the second directly and the first only slowly and indirectly. That asymmetry shapes where the effort goes.

On the retrieval side, the inputs are ones you recognise from search. Pages that rank for the underlying questions. Content that is readable in raw HTML, since several major AI crawlers fetch JavaScript without executing it according to the best available study. Facts about your business stated explicitly rather than implied. Consistency between what your site says and what third-party sources say, because a contradiction gives the system a reason to hedge or skip you.

The strongest single lever is usually not on your site at all. A 2025 study found AI search systems favour third-party authoritative sources over brand-owned pages and social content. Being described accurately in places that are not yours does more than another page you wrote about yourself. That is why this work runs alongside digital PR rather than replacing it.

• Rankings on the underlying questions, since retrieval leans on search
• Content readable without JavaScript execution
• Explicit, consistent facts about what you do and who you serve
• Third-party coverage, which the evidence suggests carries more weight than owned content
• Contradictions across sources removed, because they make a brand unsafe to name

Correcting what assistants say about you

A separate problem from being absent is being described wrongly. Assistants confidently state outdated pricing, name services you dropped years ago, attribute the wrong location, or confuse you with a similarly named business. Buyers act on those descriptions without ever visiting your site.

The fix is not a request to the model provider, because that mostly does not exist as a service. It is upstream. Find where the wrong fact lives, which is usually an old page of yours, a stale directory record, a press mention from a previous positioning, or a third-party roundup nobody has updated. Correct or supersede those sources, and make the current fact easy to find and unambiguous.

This takes time and it does not always fully resolve. Some wrong information persists because it sits in training data rather than in a retrievable page. We say so rather than promising a correction we cannot force, and we focus on making the accurate version dominant in what retrieval finds.

Building the tracking into a programme

Month one is baseline and infrastructure. Agree the prompt set, agree the competitor set, run the first full measurement, and audit the input layer: crawler access, rendering, fact clarity on site, and consistency across the third-party sources that describe you. That baseline is the number everything after is judged against.

Months two onward run a loop. Fix the input problems found in the audit, in priority order. Re-run the measurement on schedule. Report share of voice by assistant, movement against competitors, and any newly surfaced misdescriptions. Feed what the answers reveal back into content planning, because the questions where a competitor is consistently named are the clearest brief you will get.

We also watch referral traffic from assistant domains in analytics, which is small in volume for most businesses but usually well qualified, since the visitor arrives having already been matched to a need. We report it as a supporting figure rather than the headline, because attributing a specific visit to a specific citation is not something anyone can do cleanly.

• Baseline measurement plus an input-layer audit in month one
• Scheduled re-runs with results aggregated across runs and assistants
• Competitor share reported alongside yours, so movement has context
• Misdescriptions logged and traced to their upstream source
• Assistant referral traffic tracked as a supporting signal, not a headline metric

What we will not claim

Nobody can guarantee a citation. Citation is a long tail across an enormous number of domains, the systems change without notice, and no agency can attribute a specific mention to a specific piece of its own work. Any proposal that promises you will be recommended by a named assistant is promising something outside the control of whoever wrote it.

We also will not sell you the tactics that circulate without evidence behind them. Publishing an llms.txt file does not improve AI visibility, and Google has said directly that no AI system currently uses it. Adding schema markup does not cause AI citations, and the controlled study that tested it found no meaningful lift. Both are worth doing for other reasons or not at all, and we will tell you which.

What we do commit to is improving every input that is genuinely within reach, measuring the outcome with a method that survives the variance, and reporting the result plainly including when it has not moved. That is a slower pitch than a guarantee. It is the one that holds up after six months.