Citations in LLMs: measure them and improve them

The practical problem is that there is no “Search Console for generative engines” yet: no tool returns, completely and reliably, how many times a site was cited and for which questions. That is why you need a structured, repeatable method that is honest about its limits, combining manual tests on a fixed sample of questions with indirect indicators drawn from the data you already have. I treat it as the measurable part of Generative Engine Optimization.

Why monitor citations

citations in LLMs are the references to a site that appear in answers generated by the large language models behind ChatGPT, Google AI Overviews and Gemini, Perplexity and Copilot. Monitoring them means knowing, for a set of relevant questions, whether your site is named as a source, at what point in the answer, and whether the information reported is correct. Improving them means acting on the factors the models assess: recognizable entity, extractable content, complete topic coverage, consistency with other sources, verifiability.

Without measurement, Generative Engine Optimization becomes an act of faith: you make sensible interventions but don’t know whether they are working. Citation monitoring serves three purposes. First, to take a snapshot of the starting point: who is cited today in the industry, with what content, and where the site is absent. Second, to set priorities: the questions where competitors are cited and the site is not are the most concrete goals. Third, to verify progress: by comparing tests over time you see whether rewritten pages start being selected.

A further reason is risk control: monitoring intercepts wrong citations — outdated data, mistaken attributions, misleading summaries — before they consolidate, and lets you intervene by publishing the correct version so that it becomes the reference source.

How monitoring works, step by step

What to record for each question

Indirect indicators

Alongside direct tests, some data you already have tells a trend. The referral traffic from the domains of engines that link sources (Perplexity first, but also ChatGPT and Copilot) is the most concrete indicator and is visible in Google Analytics. The growth in direct searches for the brand name in Google Search Console signals that more people encounter the brand and then search for it: often it is the effect of citations that don’t produce immediate clicks. Thetrend of impressions on queries phrased as questions indicates whether content rewritten in answer format is gaining ground. None of these data isolates citations, but together they give a reliable direction.

How to act on the monitoring results

Each gap suggests a different intervention.

The limits of citation monitoring

It is right to be transparent about the limits. The answers of generative engines are not deterministic: the same question can give different results hours apart, for different users, in different geographic areas. The sample of questions is necessarily partial with respect to all possible requests. And engines change behavior without notice. That is why monitoring has to be read as a trend, not as an exact metric: it serves to understand whether you are heading in the right direction, not to produce a number to put in a report as if it were the average ranking. This honesty about the method is part of the work.

Supporting tools

Tools are emerging that automate part of these tests — they query the engines on lists of prompts and aggregate the citations — and they can save time on large samples. They must be used with the same caution: their data are estimates, must be checked manually on a sample and interpreted in light of the non-determinism of answers. They remain a support for analysis, not a source of truth. I integrate them where they help, keeping manual control over the most strategic questions.

Citations, authority and entities

The factor that, more than any other, determines whether a site is cited is how consistently the web as a whole associates it with a topic. This is built with a well-defined entity (schema Organization and Person, verified profiles, aligned information), with mentions and links from relevant sources — the field of digital PR and link building — and with complete coverage of the topic on the site, in coordination with content creation. Monitoring measures the effect; the entity’s authority is the cause.

When the citation is wrong

The most delicate case is not absence from the answer, but the citation that reports wrong information about the brand: a figure no longer valid, a wrong attribution, a distorting summary. There is no universal channel to correct every single answer of every engine. The strategy that works is indirect and on two levels. On your own site: publish the correct version explicitly, with date and source, structured as a direct answer to the question, so that it becomes the easiest and safest source to use. On external sources: identify the sites spreading the wrong information and ask for a correction, because that is often where the model learned it. Over time, if the correct version is clearer and more repeated, the engines tend to align.

How monitoring fits into SEO work

Citation monitoring is not an isolated activity: it plugs into the SEO work already under way. The questions in the set come from the same keyword research that guides the content; the interventions that emerge — rewriting blocks, adding sections, updating pages — are the same as in content creation andFAQ optimization; strengthening the entity and mentions goes through structured data and digital PR. Monitoring adds a lens: it shows which of these interventions, and on which topics, are moving the citations, and therefore where to concentrate energy in the next cycle.

How to measure AI visibility and share of citations

Measuring a brand’s presence in AI answer engines requires a stable sample of questions and readable indicators. There is no single permanent ranking: answers can change depending on the prompt, the language, the context and the sources the model considers most reliable. That is why AI citation monitoring must compare repeated surveys, not a single answer. You start from informational, commercial and comparison queries that represent the audience’s different intents. For each question I record whether the brand is cited, which page or source is attributed, in what position it appears and whether the description is correct.

From competitor monitoring to interventions

Comparison with competitors is not meant to copy their texts or metadata. It serves to understand which signals make a source more citable: clarity of the offer, verifiable data, organization of content, internal links and consistency of digital identity. From there I define a concrete SEO strategy and AEO/GEO: fixing the pages that generate ambiguity, strengthening entities, making the relationships between services and use cases explicit and updating the sources that the public and language models can verify.

How I turn data into decisions

Each client dashboard collects questions, survey date, engine used, citations, sources, competitors and recommended actions. The value is not an isolated score, but the sequence of observation, hypothesis, intervention and new check. In this way the analysis becomes part of ongoing SEO work and lets you measure whether a page is more visible, more correct and more useful in AI answers.

Frequently asked questions about citations in LLMs

Request citation monitoring

I define the set of questions relevant to your business, measure how generative engines cite you today and set up the interventions to increase correct citations, with periodic checks.