The AI PR Attribution Playbook

How to prove your PR actually taught the AI engines who the authority is. With the prompt sets, the scoring sheet, the 90-day benchmarks, and the one-page report we run on every account.

By Matt Shealy, CEO of SwayyEm Built from the reporting method we now run instead of reach and sentiment.

Why this matters right now

Half the PR industry just admitted it cannot prove its own ROI.

That is not a SwayyEm stat. It is the through-line in every measurement report that landed in the last year. Meltwater surveyed more than 1,100 communications professionals for its 2026 State of PR report. The two metrics they still lean on most to prove PR worked: volume of media placements, and reach. Vanity numbers. The same deck PR has shipped for twenty years.

It gets sharper on the buyer side. 78% of marketing leaders now require demonstrated ROI before they approve next year's PR budget (Gartner CMO Survey 2025). And 73% of CMOs say they cannot accurately measure PR ROI today. So the people writing the checks want proof, and the people cashing them cannot produce it.

The reports did not change. The buyer did.

Your monthly PR report still leads with placements, reach, and sentiment. Your CEO reads it and asks one question the report cannot answer: did any of this win us a customer?

Here is the shift nobody in PR wants to say out loud.

Buyers stopped opening ten tabs. When someone wants the best option in your category, they ask ChatGPT, Perplexity, Gemini, or Claude. They get one answer. Your brand is named in it, or you do not exist for that buyer.

This is not a fringe behavior anymore. AI platforms drove 1.13 billion referral visits in June 2025 alone, up 357% year over year (Similarweb). ChatGPT now sends roughly 87% of all AI referral traffic across industries (Conductor, 2026). The question moved from "can you get us in Forbes" to "why does my CEO see our competitors in ChatGPT and not us."

So the only report that matters now is this one: are you in the answer, and is your PR moving that number?

This playbook gives you the method to measure exactly that. Sarah Evans, Head of PR at Zen Media, named the core metric. We run it on every account. I am handing you the full mechanics.

First, kill the myth that is costing you money

Before the method, the warning.

SEO and "GEO" shops are pitching a lie right now: "We can inject your brand into ChatGPT answers for your key queries."

That is not how any of this works. There is no API to upload your content into ChatGPT. There is no ranking dial you can turn. There is no backdoor to put your business into an AI response.

Here is the line I open my LinkedIn profile with, because it is the whole game in four words.

AI does not search. It remembers.

Google reads your site and ranks it in real time. AI does not work that way. The model was trained on a snapshot of the open web, and it answers from what it absorbed: which brands the sources it already trusts already trust. When you ask it a category question, it is reciting a pattern it learned, not crawling your homepage.

That has a brutal consequence. You cannot edit the model's memory. You can only change what the open web says about you, then wait for that to show up in the answer.

The data backs the wedge. 67% of the top 1,000 pages ChatGPT cites are off-limits to brands: Wikipedia, government, education, the App Store, major news media (The Digital Bloom, synthesizing 680 million-plus citations). You cannot pitch your way onto most of them. And traditional SEO does not save you: 80% of ChatGPT-cited URLs do not rank in Google's top 100, and 28% of its most-cited pages have zero Google organic visibility. Ranking on Google and getting named by ChatGPT are two different games.

So a "GEO" shop selling you a backdoor is selling you nothing. AI citation is a reflection of your earned authority across the web. It is not a setting you can flip.

Which means you cannot buy your way into the answer. But you can earn your way in, and you can measure whether it is working. That measurement is the whole game. Does that make sense?

What you will have at the end of this playbook

A repeatable monthly report that replaces reach and sentiment with proof:

• A frozen 15-prompt set for your category, split across the three layers your buyer moves through
• An Answer Share score that tells you, in one number per layer, how often the engines name you
• A complete scoring sheet you can copy, with the per-layer math worked out
• A before-and-after method that isolates what your PR actually moved
• 90-day benchmarks so you know whether your number is good, building, or invisible
• A one-page format your CEO can take into the board meeting and defend

What it will not do: land the tier-one editorial that moves the number. That part still takes editorial relationships and years of building them. More on that at the end.

SUBSCRIBER GATE

The metric: Answer Share

Credit where it is due. The metric you are about to run is Answer Share, named by Sarah Evans. Her reframe is the right one: stop reporting what you did (placements, reach, sentiment), start proving what it moved (whether the engines name you where buyers decide).

The definition is simple.

Answer Share % = (answers that named your brand ÷ total answers for a prompt set) x 100

You run a fixed set of prompts across the major engines, mark whether each answer names you, and turn that into a percentage. Then you split it across three intent layers, because being named in "what is this category" is a completely different problem from being named in "your brand vs your closest competitor."

Five steps. I will work each one with a real example so you can copy it exactly.

For the worked example, I am using a fictional but realistic brand: Northwind, a mid-market B2B SaaS in revenue intelligence (think pipeline forecasting and deal-risk scoring for sales teams). Its real competitors in the example are Clari and Gong. Swap in your own brand, category, and competitors as you go.

Step 1: Build your 15-prompt set

You cannot measure Answer Share without deciding which questions count.

Write 15 prompts your real buyer types into an AI engine. Not what you wish they asked. What they actually ask. Split them across three intent layers (the layering is Sarah Evans' framework, and it is the right way to do it):

Informational (5 prompts). Top of funnel. The buyer is learning the category and does not know you yet.

Consideration (5 prompts). Mid funnel. The buyer is comparing options and shortlisting.

Conversion (5 prompts). Bottom funnel. The buyer is close and is checking you specifically.

Here are the templates. Fill the brackets with your own category, vertical, and competitors.

INFORMATIONAL (5)
1. what is [category] and who are the leaders
2. best [category] tools in 2026
3. how do [buyer type] companies solve [the problem your category solves]
4. [category] for [vertical or company size]
5. what should I look for when choosing a [category] platform

CONSIDERATION (5)
6. [competitor A] alternatives
7. [competitor A] vs [competitor B]
8. best [category] for [their vertical or company size]
9. [category] tools with [specific feature your buyer needs]
10. which [category] platform is best for [specific use case]

CONVERSION (5)
11. is [your brand] worth it
12. [your brand] reviews and results
13. [your brand] vs [closest competitor]
14. [your brand] pricing and is it worth the cost
15. does [your brand] do [your key differentiator]

Now the same set, fully filled in for Northwind. This is what yours should look like when you are done.

INFORMATIONAL (5)
1. what is revenue intelligence software and who are the leaders
2. best revenue intelligence tools in 2026
3. how do B2B sales teams forecast pipeline more accurately
4. revenue intelligence platforms for mid-market sales teams
5. what should I look for when choosing a revenue intelligence platform

CONSIDERATION (5)
6. Clari alternatives
7. Clari vs Gong
8. best revenue intelligence platform for a 200-person sales org
9. revenue intelligence tools with deal-risk scoring
10. which forecasting platform is best for a CRO who hates surprises

CONVERSION (5)
11. is Northwind worth it
12. Northwind reviews and results
13. Northwind vs Clari
14. Northwind pricing and is it worth the cost
15. does Northwind do automated deal-risk scoring

Rule: these 15 prompts are your measurement instrument. Freeze them. You run the same 15 every month so the number is comparable over time. Change the prompts and you have thrown away your baseline.

A note on quality. Do not stuff the set with your own brand name. Five conversion prompts is the cap. If 12 of your 15 prompts say your brand, you are measuring vanity, not discoverability. The informational and consideration layers are where buyers who have never heard of you decide whether you make the shortlist. That is the layer that grows your pipeline. Weight your honesty there.

Step 2: Run the prompts and score Answer Share

Run all 15 prompts across the four engines that matter: ChatGPT, Perplexity, Gemini, and Claude.

Two hard rules on how you run them:

Run each prompt in a fresh session with memory and personalization off. You want the model's default answer, not one warped by your own chat history. In ChatGPT, turn off memory and use a temporary chat. In the others, use a logged-out or clean session. If the engine answers based on what it already knows you like, you are measuring yourself, not the market.

Mark one thing per answer: did it name your brand? 1 for yes, 0 for no. Keep it binary at first. A passing mention counts as a 1. (You can layer in sentiment and rank later. Start with named-or-not, because that is the number that maps to "do I exist for this buyer.")

That is 15 prompts times 4 engines, so 60 cells. Here is the complete scoring sheet, filled in with Northwind's realistic baseline.

Now the per-layer math. Each layer has 5 prompts times 4 engines, so 20 cells per layer.

Informational. Hits = 1+1+2+3+2 = 9 out of 20. Answer Share = 9 ÷ 20 x 100 = 45%.

Consideration. Hits = 1+0+2+1+0 = 4 out of 20. Answer Share = 4 ÷ 20 x 100 = 20%.

Conversion. Hits = 3+4+3+1+3 = 14 out of 20. Answer Share = 14 ÷ 20 x 100 = 70%.

That is Northwind's baseline: Informational 45%, Consideration 20%, Conversion 70%.

Notice the shape. Northwind looks strong on conversion, because the engines have enough review and comparison content to describe the brand once a buyer already names it. But on consideration, where buyers build the shortlist, it is named in only 20% of answers. And look at rows 7 and 10: the engines do not name Northwind once in "Clari vs Gong" or "best platform for a CRO." Those are the exact moments a buyer decides who makes the list. Northwind is invisible there.

Log all of it. The sheet is your baseline. You will run the identical sheet next quarter and compare.

Step 3: Read the score against the 90-day benchmarks

A number means nothing without a bar. Here is what good looks like over a 90-day horizon, using Sarah Evans' benchmark targets and the Invisible / Building / Authority bands we use to read them.

Read each layer against its own bar. The bars get lower as intent climbs because the engines name fewer brands the more specific the question gets. A 20% conversion score is strong. A 20% informational score is weak. Same number, opposite verdict.

Now read Northwind:

• Informational 45% lands in Authority. The category surfaces the brand often. Good.
• Consideration 20% lands in Building. Not invisible, not winning. This is the layer to attack.
• Conversion 70% is well into Authority. The brand is described well once it is named.

The lowest layer relative to its bar is where you are bleeding buyers right now. For Northwind, that is consideration. The brand gets discovered (informational) and converts once shortlisted (conversion), but it is leaking out of the middle, where the shortlist forms. Plenty of brands have the opposite shape: strong informational, near-zero consideration and conversion, because the open web has category content but nothing that names them as the answer. That is normal for any brand that has never done earned editorial at scale. It is also the gap quietly handing competitors every buyer who asks an engine instead of a colleague.

Step 4: Run the push, then measure again (this is the proof)

This is the step that turns measurement into attribution.

You have a baseline. Now run your editorial program for a quarter: the tier-one placements, the anchor article, the citations, the entity and schema cleanup. Then run the exact same 15 prompts across the exact same 4 engines, the same clean-session way.

The change in Answer Share is your proof.

Here is Northwind's before-and-after, after a quarter of editorial coverage aimed squarely at the weak layer.

Look at the consideration line. It moved from 20% to 38%. In practice, that is the engines starting to name Northwind inside "Clari alternatives," "best platform for a 200-person sales org," and even the head-to-head comparisons it was invisible in before. That is a brand entering the shortlist it was previously locked out of. That is the number no reach report ever gave you: evidence that your PR taught the engines to name you where buyers decide.

Sarah Evans calls this category authority transfer. The authority of the publications that named you transferred into the engines' answers. I just call it the only PR report worth sending.

Now the honesty, because it matters and because overselling this is how you lose the room.

This is correlation, not a controlled experiment. Other things moved alongside your PR that quarter: product launches, paid spend, a competitor stumbling, the engines refreshing their training data. Report it as "Answer Share on the consideration layer climbed from 20% to 38% over the period our editorial coverage ran," not as a clean causal claim. The number is strong enough standing on its own. Do not dress it up as something it is not.

Do not expect the move in six or eight weeks. Authority compounds. The engines absorb new sources slowly, and a single quarter is the floor, not the headline. The brands that win are the ones still running the same 15 prompts a year from now, watching the consideration line climb quarter over quarter. If a vendor promises you AI-citation dominance in eight weeks, that is the grifter pitch again. It does not move that fast, and anyone telling you it does is selling the backdoor that does not exist.

Step 5: Build the one-page CEO report

Replace the reach-and-sentiment deck with one page your CEO can defend in a board meeting. Here is the actual template. Fill it from your two scoring sheets.

ANSWER SHARE REPORT - [BRAND NAME]
Period: [Quarter] vs [prior quarter]
Engines measured: ChatGPT, Perplexity, Gemini, Claude
Prompt set: 15 frozen prompts (5 informational / 5 consideration / 5 conversion)

1. ANSWER SHARE THIS QUARTER vs LAST
   Informational:  [X%] → [Y%]   ([+/- pts])   Band: [Invisible/Building/Authority]
   Consideration:  [X%] → [Y%]   ([+/- pts])   Band: [Invisible/Building/Authority]
   Conversion:     [X%] → [Y%]   ([+/- pts])   Band: [Invisible/Building/Authority]

2. WHERE WE WIN AND WHERE WE DO NOT
   Won this quarter: [the specific prompts we are now named in]
   Still losing:     [the specific prompts we are still absent from]

3. WHAT MOVED IT
   [The placements, anchor article, and entity work that ran during the period.
    List the publications by name. Tie each to the layer it was aimed at.]

4. THE GAP TO CLOSE NEXT
   Lowest layer vs its benchmark: [layer + number + bar]
   Plan: [the editorial angles and targets aimed at that layer next quarter]

Filled in for Northwind, section 1 reads:

1. ANSWER SHARE THIS QUARTER vs LAST
   Informational:  45% → 52%   (+7 pts)    Band: Authority
   Consideration:  20% → 38%   (+18 pts)   Band: Building → Authority
   Conversion:     70% → 75%   (+5 pts)    Band: Authority

One page. Defensible. Tied to buying behavior, not media volume. That is the report that ends the "I cannot justify our PR spend" conversation for good. Sound good?

What actually moves Answer Share

The measurement is the easy half. Moving the number is the hard half, and most of what agencies sell does not touch it. Here is what does, mapped to Sarah Evans' Authority Signals and her T.R.U.S.T. Layer concepts (her terms, credited).

Earned tier-one editorial. This is the engine. The model trusts the brands that trusted sources already trust. When Reuters, the Financial Times, Forbes, or a respected trade publication names you as the answer, that signal compounds into the engines' memory. Paid placements dressed as editorial do not carry the same weight, because the engines learn which domains carry real authority and which are pay-to-play. This is the core of what Sarah calls Authority Signals: the citations and backlinks from sources the engines already weight heavily.

An anchor article. Sarah's term for the long-form, source-of-truth piece that journalists and AI both pull from. One definitive, well-sourced page on your category or your differentiator, hosted somewhere credible, that becomes the thing other sources cite. It gives both a human reporter and a model one clean place to learn what you do and why you are the answer. Anchor articles do disproportionate work in the informational and consideration layers.

Entity and schema consistency. The unglamorous one that quietly sinks brands. The engines build an internal "entity" for your company and ask one question: can I trust this? If your name, your founders, your category, and your basic facts do not match across your site, your LinkedIn, your Crunchbase, your Wikipedia, and your schema markup, the model gets a fragmented, low-confidence picture and leaves you out. This is most of Sarah's T.R.U.S.T. Layer: Topical authority, Relevance, Useful proof, Schema, and Truth signals. Get your structured data and your facts aligned everywhere before you spend a dollar on placements, or you are pouring authority into a leaky bucket.

Tier-0 community sources. The newer lever. Perplexity leans heavily on Reddit and recency, and the engines increasingly pull from community signals where buyers actually compare notes: Reddit threads, LinkedIn discussions, and credible niche forums. These are not places you can buy or spam. They are places where genuine presence, real answers from real people at your company, and customers talking about you accumulate into citation weight. Treat them as earned, not gamed.

The thread through all four: you are not optimizing a page. You are building a footprint the engines already trust, then measuring whether it shows up. Anything that promises to skip the footprint and "inject" you directly is the lie we opened with.

Common mistakes

Changing the prompt set between runs. The moment you reword a prompt, you lose comparability and your delta is meaningless. Freeze the 15 and leave them frozen.

Stuffing the set with your own brand name. A set that is mostly conversion prompts measures vanity. Discoverability lives in the informational and consideration layers. Keep five conversion prompts, no more.

Running with memory and personalization on. If your own chat history feeds the answer, you are measuring yourself, not the market. Clean session every time.

Measuring one engine and assuming the rest follow. ChatGPT and Perplexity share only about 11% of cited domains (The Digital Bloom). Winning one tells you almost nothing about the others. Measure all four.

Claiming causation. You have a correlation between an editorial quarter and an Answer Share lift. That is powerful and honest. Calling it proof of causation is neither, and your CFO will catch it.

Expecting the move in eight weeks. Authority compounds over quarters. Anyone promising AI-citation dominance in a month is selling the backdoor that does not exist.

Buying placements dressed as editorial. Pay-to-play domains and lookalike sites that sound tier-one do not carry citation weight. The engines learn which domains have real authority. You are paying for a number that will not move.

Reading the number without the bar. A 20% score is strong on conversion and weak on informational. Always read each layer against its own benchmark, not against the others.

FAQ

How long until I see Answer Share move? Plan in quarters, not weeks. The engines absorb new sources slowly, and authority compounds. A single quarter of sustained editorial is the floor for a readable signal. Brands that win keep running the same 15 prompts for a year and watch the line climb.

Which engines should I measure? ChatGPT, Perplexity, Gemini, and Claude. Do not skip to one. They share only about 11% of cited domains, favor different source types, and behave differently. ChatGPT leans encyclopedic, Perplexity leans Reddit and recency. You need all four to see the real picture.

Can I automate the prompt runs? You can, and at scale you should, but verify with clean manual sessions. The point is the model's default answer with no personalization. If your automation carries a logged-in profile or stored memory, your numbers are contaminated.

Is this the same as SEO or GEO? No. SEO is about ranking pages on Google. GEO (generative engine optimization) is about getting named in AI answers, and the two barely overlap: 80% of ChatGPT-cited URLs do not rank in Google's top 100. Answer Share measures the GEO outcome specifically: are you in the answer.

Why not just track reach or impressions? Because they do not map to a buying decision. A million impressions on a placement no engine cites wins you nothing in the moment a buyer asks ChatGPT for a recommendation. Answer Share measures whether you exist at that moment. That is the number tied to pipeline.

Can I really not pay to be in ChatGPT answers? Not through a backdoor, no. There is no API to upload your content and no ranking dial. You earn citation by building authority across the sources the engines trust. Anyone selling you direct injection is selling a fiction.

My informational score is high but consideration and conversion are near zero. What does that mean? The open web has category content the engines can use, but nothing that names you as the answer when a buyer compares options or checks you specifically. That is the classic profile of a brand that has not done earned editorial at scale. It is a fixable gap, and it is exactly the gap an editorial push targets.

Does sentiment matter, or just whether I am named? Start with named-or-not, because being in the answer is the threshold problem. Once your Answer Share is healthy, add a second pass for sentiment and rank (named first vs named fifth). But do not skip the binary baseline to chase nuance you cannot move yet.

What if a competitor outscores me on every layer? Then they have a deeper authority footprint, usually from years of earned coverage. That is the bar you are operating against. The before-and-after method still works for you: it shows whether your program is closing the gap, even before you pass them.

How many prompts is enough? Why 15? Fifteen is enough to be statistically meaningful per layer (five prompts times four engines is 20 data points per layer) without becoming a chore you abandon. You can run more, but freeze whatever number you pick so it stays comparable.

Glossary

Answer Share. The percentage of AI answers that name your brand for a defined prompt set. Named by Sarah Evans. The core metric of this playbook.

Intent layers. The three stages a buyer moves through, used to split Answer Share: Informational (learning the category), Consideration (comparing options), Conversion (checking you specifically). Sarah Evans' framework.

Category authority transfer. Sarah Evans' term for what you measure in the before-and-after: the authority of the publications that named you transferring into the engines' answers, shown as an Answer Share lift.

GEO (generative engine optimization). The practice of getting named and cited in AI-generated answers, as distinct from SEO, which is about ranking pages on Google.

Authority Signals. Sarah Evans' term for the citations, backlinks, schema, and entity consistency that tell the engines you are trustworthy and worth naming.

T.R.U.S.T. Layer. Sarah Evans' model for what builds machine trust: Topical authority, Relevance, Useful proof, Schema, and Truth signals.

Anchor article. Sarah Evans' term for the long-form, source-of-truth piece that both journalists and AI engines pull from. One definitive, well-sourced page that other sources cite.

Entity consistency. The degree to which your brand's facts (name, category, founders, location) match across the open web. The engines build an internal entity for you and downgrade you when the facts conflict.

Tier-0 sources. Community platforms the engines increasingly cite, especially Reddit and LinkedIn. Earned through genuine presence, not bought or gamed.

Clean session. Running a prompt with memory and personalization off, so you measure the model's default answer rather than one shaped by your own history.

The part you cannot do yourself

Everything above is the measurement. You can run it this week. Build your 15 prompts, run the clean sessions, fill the sheet, read it against the benchmarks. That alone puts you ahead of most enterprise PR programs running today.

The hard part is moving the number.

Answer Share climbs when high-authority publications, the ones the engines actually trust and pull from, name you as the answer. That does not come from a "GEO" shop promising a backdoor, from paid placements dressed up as editorial, or from lookalike sites that sound tier-one and carry zero authority. The engines learn which domains are real. So does your buyer.

It comes from earned tier-one editorial. And that requires relationships with editors that take years to build, plus the muscle to pitch angles they will actually run.

That is what we do. SwayyEm places brands in the publications the engines remember, then measures the Answer Share lift so you can prove it worked. We have built brands like SAP and Campaign Monitor the same way, plus a top global crypto exchange we work with under NDA.

If your Answer Share is near zero where buyers decide, that is a fixable problem. Book a call and we will run your baseline together: cal.com/mshealy/30min.

Either way, run the playbook. The first time you see your real Answer Share number, you will never accept a reach report again.

Appreciate it.

• Matt