Public relations has always had a measurement problem, but it used to be an argument about attribution. You could at least point at the coverage. In 2026 the problem is different and worse: a growing share of the audience never sees the coverage at all, because an assistant read it and summarised it for them.
That is not a prediction. It is visible in Search Console today, and once you know what to look for it is unmistakable. The backdrop is not subtle either: organic search clicks have fallen roughly 42% since AI Overviews began expanding, according to Define Media Group research reported by Search Engine Land — a decline that began at around 16% and deepened as coverage grew.
The reporting gap, measured
We run search and answer-engine work for a US press-release distribution platform. The client is not named here under NDA, but the pattern in its data is worth publishing because we have not seen it described anywhere with numbers attached.
Across six months of Search Console data we found roughly sixty conversational, natural-language queries sitting at an average position of about 4.4 — page one, frequently top five. These are not head terms. They are full sentences, the kind of thing a person types into an assistant rather than a search box:
- which press release distribution companies have the strongest analytics for tracking backlinks, opens, and media pickup
- is business wire actually the best press release distribution service or are there better options now
- what is the best press release distribution solution under a fixed budget for investor news
Every one of those is a buying question asked at the decision stage. The platform ranks on page one for all of them. They produced almost no clicks.
Separately, eight questions about the brand itself — who it is, what it costs, which industries it serves, what its success rates are — sat at an average position of roughly 1.1. Answer engines have resolved the company as an entity and are answering questions about it directly, using its own pages as the source.
And one more thing, which we did not expect. A raw large-language-model system prompt appeared in the query report as a ranked query. Not a person's question — an instruction to a model, logged as search demand. Machine traffic is now measurable in the same report you send your client.
The blunt version: a single top-three ranking on an informational term generated more than 40,000 impressions and zero clicks in that window. Ranking and traffic have come apart.
Why the old PR metrics stopped working
Short answer: pickup counts, wire backlinks and referral traffic now measure the wrong thing. Google treats optimised links in syndicated releases as link spam, and most AI citations produce a read without a click. Earned editorial coverage from a named outlet is the one traditional metric that survives intact.
Pickup count. A wire release syndicates to hundreds of mirror domains. That number has never correlated with outcomes, and it correlates even less now. The wire services have moved on themselves — PR Newswire now sells an AI-optimised release format explicitly aimed at becoming a cited source inside ChatGPT and Gemini, which tells you where the category thinks the value has gone.
Backlinks. Google's spam policies name the practice directly, listing "links with optimized anchor text in articles, guest posts, or press releases distributed on other sites" as link spam. Earned editorial links still matter. Wire links are not that.
Referral traffic. Still real, still worth tracking, but it now systematically understates reach — because the most common outcome of a citation is that someone reads the answer and never clicks. Treating assistant referrals as the measure of assistant influence will tell you the channel is negligible when it is not.
What you can actually measure
Short answer: four methods are available — assistant referral traffic, conversational queries in Search Console, a tracked prompt set, and entity resolution checks. None is complete on its own. Run all four, report the prompt set as a trend rather than a rank, and state plainly what each method cannot see.
| Method | What it sees | What it misses | Effort |
|---|---|---|---|
| Assistant referral traffic | Real sessions from ChatGPT, Perplexity and similar, in GA4 | Every citation that was read but not clicked, which is most of them | Low |
| Search Console conversational queries | Long natural-language queries and the positions you hold on them | Whether an assistant actually used you; attribution to a specific answer | Low |
| Tracked prompt set | Whether you are named, how you are described, which source is cited | Personalisation, regional variation, and day-to-day model drift | Medium |
| Entity resolution checks | Whether the model has your identity, pricing and category correct | Nothing about volume or demand | Low |
Assistant referral traffic
Start here because it takes an afternoon, but hold it loosely. Segment assistant referrers in GA4 as their own channel rather than letting them fall into direct or generic referral. You will get a small number. The number is real. It is a floor, not a measure.
Conversational queries in Search Console
This is the underused one. Filter your query report for queries over roughly eight words, or containing question words, and look at the positions. Most PR and communications teams have never run that filter and are surprised by what comes back — page-one positions on precisely the questions their buyers ask, invisible in every report because the clicks are near zero.
That filter is the cheapest diagnostic in this whole article. Run it before you spend anything.
A tracked prompt set
This is the closest thing to a rank report that exists for answer engines, and it has to be built by hand.
- Write thirty to fifty questions your buyers actually ask. Do not invent them. Take them from sales calls, from your Search Console conversational queries, and from the phrasing competitors use in their comparison pages.
- Fix the wording and freeze it. The prompt set only produces a trend if the prompts do not change. Version it.
- Run it on a schedule across the assistants that matter to your market — typically ChatGPT, Perplexity, Google AI Mode and Copilot — on the same cadence every time. Monthly is enough to start.
- Record four things per prompt: were you named, in what position within the answer, how were you described, and which URL was cited.
- Track the description, not just the mention. Being named with the wrong price or the wrong category is a problem that a mention count will hide.
Entity resolution
Ask the assistants directly who you are, what you sell, what you cost and who you serve. If the answers are wrong or vague, no amount of content will fix it until the entity is clean: consistent naming, a real pricing page, schema that marks up the organisation and its people, and profiles that agree with each other. Our guide to getting cited by AI assistants covers what the evidence does and does not support here, and the llms.txt guide covers where that file genuinely helps and where it is oversold.
What the research says moves citation likelihood
Measurement tells you where you stand. If you then want to move it, the most useful controlled evidence available is still GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024), which tested nine optimisation tactics against a 10,000-query benchmark.
The three strongest were adding statistics, citing sources, and adding quotations, which improved visibility by roughly 30 to 40 percent on the paper's position-adjusted metric. Keyword stuffing performed worse than doing nothing at all.
Two things follow for PR specifically. First, the release itself should carry sourced numbers and named quotes rather than adjectives — which is what a good release did anyway, and is now measurable rather than merely tasteful. Second, lower-authority domains gained the most from these tactics in the study, so this is one of the few areas where a smaller brand is not automatically outspent by a larger one.
The split that matters more than the total
There is a second division that belongs in the same report, and it catches people out because it looks like growth.
Informational PR content — what a press release is, how to write one, what a communications team does — attracts a genuinely global audience. Very little of that audience can buy a US distribution service. In the platform's data, the United States and one South Asian market produced almost the same number of clicks over six months. The United States supplied more than four times as many impressions to get there, at roughly a quarter of the click-through rate. The other market converted better on paper and could not buy the product.
Aggregate that into one line on a report and it reads as healthy growth in both traffic and engagement. Split it by market and by intent and it reads as a content engine pointed at an audience that was never going to convert.
So run three splits, not one:
- Brand versus non-brand. If brand is carrying the clicks, the campaign is being credited with demand it did not create. In the account above, brand terms were around one percent of impressions and the clear majority of clicks.
- Commercial versus informational. Report the commercial line as the headline number.
- Buying market versus everything else. Define the list before you look at the data, not after.
None of this requires new tooling. It is the same export, cut three ways, and it changes what the next quarter of work gets pointed at.
What a defensible baseline report looks like
Three sections, and the discipline is in keeping them separate.
Earned coverage. Named outlets, with links, separated from syndicated mirrors. One line each. This is the part your client already understands and it should stay.
Search. Split into brand, commercial non-brand, and informational. Report the commercial non-brand line as the headline. In the account described above, blended reporting made the last six months look like a massive win; the commercial line had barely moved. Splitting the report was the single most useful change we made to it.
Answer engines. Prompt set results as a trend, plus referral sessions as a floor, plus the entity check. Say plainly that this is a sample, not a census.
What not to promise
There is no citation rank tracker, and anyone selling one is selling a sampled prompt set with a dashboard on it. Assistant answers vary by user, by region, by session and by model version, so two people running the same prompt on the same day can get different answers. A trend across a fixed prompt set run repeatedly is meaningful. A single screenshot is not.
Be equally careful about causation. You can show that a page was cited. Showing that a release caused the citation requires the same before-and-after discipline as any other channel, and usually you will not have a clean control.
The right posture with a client is the one that survives contact with a sceptical CFO: here is what we measured, here is the method, here is what the method cannot see.
Where to start
If you do one thing this week, run the conversational-query filter on your own Search Console. If you rank on page one for sentence-length buying questions and earn nothing from them, you have the same gap the platform above had, and you now know it exists — which is more than most of the market can say.
We build this measurement layer for PR agencies, communications teams and distribution platforms, usually alongside the search and answer-engine work and the PR and brand building it reports on. If you want the prompt-set template we use, get in touch.
