Can you measure whether your earned media is what AI models cite about you?

The question, verbatim

This arrived in our Search Console as a single query, 23 words long:

i'm a paid media director. is there a way to measure whether our earned media is what ai models cite about our candidate?

It is a better question than most of the answers currently available to it. It separates two things the industry routinely merges: whether AI cites you at all, and whether the thing being cited is the work you paid for.

We are going to answer it properly, including the parts where the honest answer is "only up to a point".

What is being sold

Search this question and you will find a fairly uniform answer, largely from one vendor's content operation, proposing a three-stage chain:

• Earned media placement.
• AI citation, measured as share of voice.
• Referred traffic, tracked through analytics and UTM parameters.

The accompanying advice is to capture a baseline, log the placement, re-audit on a cadence, and attribute citation shifts to the media activity. Some of that is genuinely good practice. The last step is where it breaks.

Three questions, three standards

First, separate three questions that get merged into one. You can answer the first two.

Most reporting collapses these three. Observing a displayed URL is straightforward and worth doing. Proving incremental influence is a different standard, and that is the one the three-stage chain quietly claims to meet.

Causal attribution requires that the effect could only have come from the cause. Almost nothing about AI citation satisfies that.

• The same facts exist in many places. If a trade outlet publishes that your organisation does X in Y, that claim probably also appears on your own site, a directory, a press release, and three aggregators that scraped the first one. A model citing any of them is citing the fact, not your placement.
• Retrieval is opaque. You see a displayed source. You do not see what was retrieved, considered and discarded, or what the model already held without retrieval.
• Outputs are not stable. The same prompt returns different sources across runs. A citation appearing after a placement may be run-to-run variance, which is why any baseline has to be repeated rather than sampled once.
• Timing is not causation. Coverage usually clusters. A placement, a press release, a LinkedIn post and a competitor's mention often land in the same fortnight.

So "citations rose after the placement" is a correlation with a sample size of one, in a non-deterministic system, with an uncontrolled confounder. That is not an attribution. It is a coincidence with a chart.

The tracer fact

There is one design that gets closer, and it is borrowed from how people test whether anything propagates: plant something that could only have come from one place.

A tracer fact is a true, specific, verifiable detail about your organisation that appears in exactly one public source and nowhere else. Not a fabrication. A real fact you simply have not published anywhere before.

Both readings are weaker than they look, and the negative is the weaker of the two.

• If the fact never surfaces, that is inconclusive. The source may have been retrieved and not selected for the answer, paraphrased past recognition, or simply not relevant to the questions asked. Nonappearance means not observed in these outputs, not unused.
• If the fact does surface, something propagated, but not necessarily your placement. Syndication, aggregation or a third party repeating the detail all carry it too.

A tracer is an optional research aid with a narrow reading, not a reliable attribution mechanism. It is more honest than the chain being sold, which is a low bar rather than a solution.

We have not run this at scale and we are not presenting it as validated. It is a design that answers the question asked, which is more than the three-stage chain does.

The research nobody in marketing reads

The most rigorous public work on what AI systems say about named entities is not in marketing. It is in election research, and it is funded by organisations with no product to sell.

The States United Democracy Center tested ChatGPT and Google AI Overview against eight standardised question types about registration, voting dates, polling places, mail voting, candidates and results counting. A preliminary round ran September to October 2025 across six states. A primary round ran January to February 2026 across Arizona, Michigan and Pennsylvania.

The accuracy result is the one that gets reported:

In the preliminary investigation, 6.9% and 8.2% of the responses from Google AI and ChatGPT, respectively, contained verifiable factual errors. In the primary investigation, conducted several months after the preliminary investigation, no responses across any platform conditions were coded as inaccurate.

Zero factual errors, a few months later. That is a genuine improvement and it is what most coverage led with.

The finding that matters more for anyone trying to be named is the one underneath it:

ChatGPT returned an incomplete list of declared candidates in 88.9% of the gubernatorial candidate queries, 35.7% of the attorney general queries, and 28.6% of the secretary of state queries.

Accuracy and inclusion are different failures

Read those two results together. Factual error rate went to zero. Completeness stayed broken in nearly nine out of ten gubernatorial queries.

Everything the system said was correct. It simply did not mention everyone who existed.

For a candidate, a business or any named entity, that is the important distinction. You can be entirely correctly described and still absent from the list. Chasing accuracy fixes the wrong failure if the actual problem is omission.

This is the same split we ask people to record separately as mention, recommendation, citation and accuracy. The election work is the strongest independent demonstration of it we have seen, on a population where the stakes are considerably higher than a lead form.

It also arrives without a product attached, which makes it worth more than most of the vendor research in this field.

What to do on Monday

• Establish which failure you have. Run a fixed prompt set and record, separately, whether the entity is named at all, whether it appears in list-type answers, and whether what is said is accurate. Those are three different columns and three different fixes.
• Baseline before the placement, not after. There is no back-fill. A placement with no before state cannot be evaluated at all.
• Log the source URLs, not just a score. Which domains are being cited for these questions matters more than a share-of-voice number, because it tells you where the answer is actually sourced from.
• Consider one tracer fact if you genuinely need to know whether a specific outlet is feeding the answers, and accept that a null result is the most likely and most useful outcome.
• Report inclusion rate separately from accuracy. If your reporting has one AI visibility number, it cannot distinguish the two failures above, and you will spend on the wrong one.

The broader question of whether earned media is worth buying for AI visibility at all is a different argument, and the headline statistic used to sell it does not say what it is used to say.

Limits

• The tracer method is a proposal. We have not run it at scale, a single tracer speaks to a single outlet, and a null result is inconclusive rather than evidence of non-use.
• A null tracer result rules out nothing on its own. Retrieval without selection, paraphrase and irrelevance to the questions asked all produce the same silence.
• The States United work covers United States election information on two platforms in named states. It is not a study of commercial visibility, and we are arguing that its structural finding transfers, not that its percentages do.
• Figures quoted are from the preliminary and primary investigations dated September to October 2025 and January to February 2026. These systems change quickly, and the accuracy improvement between those two rounds is itself the proof of that.
• We sell measurement work, so a piece arguing that measurement is harder than vendors claim is not disinterested either. The difference is that our conclusion reduces what we can honestly bill for.

If someone sells you attribution of an AI citation to a specific placement, ask what would have to be true for that claim to be falsifiable. The answer is usually nothing, which is the problem.