“Narrative Accuracy is the degree to which an AI assistant’s claims about your brand — its ingredients, its positioning, its price point, its persona, its provenance — are true to your owned positioning. It is the closest measurable proxy for brand value in the AI Purchase Channel — and it is what share-of-voice metrics, inherited from PR measurement, were never built to detect.” — Slingso, Narrative Accuracy vs Share of Voice in the AI Purchase Channel
A premium skincare brand we’ll call Brand B runs an AI-monitoring tool. The dashboard reports 82% share of voice across ChatGPT, Gemini, and Perplexity for “vitamin C serum.” The brand’s mention count is up 14% quarter-over-quarter. The marketing team is pleased; the dashboard says brand presence is healthy.
What the dashboard does not tell them: in 60% of those mentions, the AI describes Brand B as a “drugstore” brand. Brand B retails at $52 a bottle. Its packaging is glass. Its ingredients are clinical-grade. Its target customer is a 35–45-year-old buying considered skincare. The AI is mentioning the brand frequently and describing it as something it isn’t — and every shopper exposed to that description, whether they’re reading the AI’s words directly or letting an AI agent act on those words on their behalf, forms a mental model of Brand B that the brand’s actual marketing has spent five years trying to build the opposite of.
This is brand equity erosion at scale, in real time, invisible to the share-of-voice dashboard. It is the second-most-expensive measurement mistake in the AI Purchase Channel — second only to the visibility/ranking mistake we covered yesterday — and it is happening at most premium brands that have started monitoring AI mentions. The metric they report up celebrates the very behaviour that’s destroying the asset they’re trying to grow.
This essay is the brand-value half of the AI Purchase Channel measurement stack. Yesterday’s essay made the case for Purchase Likelihood as the revenue metric — what share of recommendations you win. Today’s makes the case for Narrative Accuracy as the brand-value metric — whether what’s being said about you, when you are mentioned, is true to who you are. They are different layers, demanding different work, measured by different instruments.
What Share of Voice actually measures
Share of voice has a specific lineage: it was a PR-era and media-measurement metric. The logic, in its native habitat, is straightforward — of all the brand mentions in this category across this set of media, what percentage names us? It treated mentions as an undifferentiated unit of presence. The implicit assumption was that mentions are roughly fungible: a mention is a mention is a mention. If you got more of them than competitors, you “owned” the conversation.
That assumption was already wobbly in PR — a feature in Vogue and a snarky paragraph in a Reddit thread are not the same mention, even if both mentioned the brand. But SoV survived because the cost of distinguishing was high (humans had to read each piece) and the volume was low enough that you could mostly trust the editorial filter. Vogue mentions were Vogue mentions; the reader knew the context.
In the AI Purchase Channel, the cost of mentions has dropped to zero. AI assistants generate mentions on the fly — improvised descriptions, paraphrased reviews, summarised positioning. Volume is enormous. The editorial filter is gone; the generative filter is in its place, and it makes things up. SoV’s foundational assumption — that mentions are roughly fungible — collapses. A “mention” can be:
- An accurate description that matches the brand’s positioning (good)
- A neutral mention that names the brand without describing it (weak)
- An inaccurate description that misrepresents what the brand is (actively harmful)
- A fabricated description containing claims the brand has never made (legally and reputationally risky)
SoV counts all four as one “mention.” It cannot distinguish them. The dashboard goes up the same whether the AI is describing Brand B accurately or calling it drugstore.
That breaks the metric.
What Narrative Accuracy actually measures
Narrative Accuracy is the metric SoV stops being. It asks a different question: when the AI assistant talks about your brand, how true is what it says?
A clean Narrative Accuracy measurement decomposes into two factors: Coverage and Correctness.
Coverage is the share of an AI assistant’s claim space about your brand that traces back to your owned content — your product pages, your schema markup, your verified reviews, your published statements. If the AI is generating descriptions of Brand B based on a retail-tier reference site that mistakenly classified it as drugstore three years ago, Coverage on that claim is zero — the brand’s owned content was not the source. High Coverage means the AI is reading the brand from authoritative anchors. Low Coverage means it’s improvising from whatever third-party scraps it has indexed.
Correctness is the share of the AI’s claims that match the brand’s actual position — verified, current, and on-brand. A claim “Brand B is a drugstore brand” has Correctness 0; a claim “Brand B is a $52 premium serum with 15% L-ascorbic acid” has Correctness 1. Correctness is per-claim, summed and weighted by claim frequency.
Narrative Accuracy at the brand level is roughly Coverage × Correctness, summarised across the AI’s full claim universe about that brand. Both factors range 0–1; the product reflects how much of the AI’s understanding of your brand is yours, accurately. A brand with Coverage 0.7 and Correctness 0.5 has Narrative Accuracy 0.35. A brand with Coverage 0.95 and Correctness 0.95 has Narrative Accuracy 0.90 — the AI is reading the brand from the brand’s own materials and getting it right.
Both factors are measurable. Coverage is measured by tracing each AI claim back to its likely source corpus — owned, third-party, or hallucinated. Correctness is measured by comparing AI claims against a brand-maintained ground-truth document (positioning, pricing, ingredients, persona). Neither is what SoV measures. SoV does not look at claim content at all.
Why the difference is brand-value-defining
Yesterday’s essay described why Purchase Likelihood is the revenue-layer metric — the AI Purchase Channel returns one or two recommendations, mention is cheap, recommendation is scarce. Narrative Accuracy is the brand-value-layer metric for a parallel reason: in this channel, the AI’s description of your brand is read more often than your own marketing’s description of your brand.
Think about the volume. A consumer who’s curious about Brand B in 2026 doesn’t necessarily land on Brand B’s website. They ask ChatGPT “is Brand B worth buying?” and read the AI’s three-sentence answer. They ask Perplexity “how does Brand B compare to Glossier?” and read the AI’s comparison paragraph. They ask Gemini “what’s the active ingredient in Brand B?” and read the AI’s claim. The AI’s description of the brand is what the buyer encounters — for many product searches, more often than the brand’s own copy. Possibly orders of magnitude more often.
If those descriptions are accurate, the AI is, functionally, a high-trust amplifier of the brand’s positioning. The buyer reads what the AI says, internalises it, and forms a mental model that aligns with how the brand wants to be seen. Brand equity compounds.
If those descriptions are inaccurate, the AI is something stranger and worse: a high-trust misamplifier. The buyer reads inaccurate claims, internalises them, and forms a mental model that the brand will spend marketing dollars trying to undo for years. Brand equity erodes — and the brand has no easy way to know, because the channel is invisible to traditional brand-tracking instruments.
SoV cannot tell you which kind of channel the AI is being for your brand. It treats the amplifier and the misamplifier identically. NA can.
NA in the delegated mode — when the AI is the reader, not just the messenger
There’s a second reason NA is load-bearing in this channel, and it is the one most operators have not adjusted for. The AI Purchase Channel has two modes of offload — assisted and delegated. In the assisted mode, the AI’s description of your brand is read by a human, who internalises it and acts on it. NA shapes the human’s mental model. That’s the case we just covered.
In the delegated mode, the shopper has handed the whole job to a personal AI agent — OpenClaw, Hermes, a custom build — and stepped back. The agent now does the comparing, the evaluating, the choosing. It is reading the AI’s claims about your brand not as persuasion, but as operating input. A claim that Brand B is “drugstore” is no longer a perception problem the brand can recover from with better copy. It is a filter. The agent screens out Brand B before the human ever hears about it, because the agent’s brief said “premium.”
Two consequences follow.
First, NA is the brand’s interface with the delegated shopper. Whatever the AI says about you is what the agent acts on. There is no second-chance moment — no landing page, no remarketing, no founder’s note in the order confirmation — where the brand gets to reset the impression. The agent acted on the claim. The transaction either happened or it didn’t.
Second, the Correctness bar is higher in the delegated mode than in the assisted mode. Humans tolerate a sloppy description and correct for it with their own taste. Agents do not. A 0.5-Correctness claim (“Brand B is a popular vitamin C serum” — true but generic) might still get the brand into a human’s consideration set. The same claim, fed to an agent whose brief is “find me a clinical-grade serum at premium price tier,” is filtered out. Sloppy NA in the assisted mode is brand erosion. Sloppy NA in the delegated mode is invisibility.
This is why NA matters now, not later. The delegated share of the channel is small today and growing every quarter. The brand that gets NA right in 2026 is also the brand that gets bought by the delegated shoppers of 2027.
The training-data feedback loop is the worst part
Here is the mechanism that turns this from a measurement problem into a strategic emergency: AI assistants do not just generate descriptions in real time. The descriptions they generate get cached, shared, screenshotted, posted in forums, indexed by search engines, and eventually fold back into training corpora. Every inaccurate AI description today is a higher-base-rate AI description tomorrow.
If ChatGPT calls Brand B “drugstore” in 60% of mentions today, those answers get screenshotted into Reddit threads, paraphrased in blog posts, indexed by Google. Those derivative pieces become inputs to the next training pass. Brand B’s “drugstore” designation calcifies. Six months later, the rate is 75%. A year later, it is 85%. The AI’s description of the brand has now drifted to a position that bears almost no relation to the brand’s actual positioning, and the brand’s marketing budget cannot outpace it.
This is brand drift in a feedback loop. SoV does not see drift; it counts mentions. NA sees drift directly: Coverage and Correctness both decline as third-party derivative content overwhelms the brand’s owned content in the AI’s reference corpus.
For premium brands — ones whose positioning is the asset — this is existential. A drugstore brand miscalled premium recovers easily; the misnomer is upmarket and the customer base is unbothered. A premium brand miscalled drugstore loses pricing power, retailer leverage, and the brand-aspirational pull that justified the price. There is no easy reset.
Five ways SoV deceives operators on brand value
Five recurring failure modes when brand teams use SoV as the primary brand-tracking metric in the AI Purchase Channel.
1. Misrepresentation gets credited as presence. “Brand B is a drugstore brand” counts the same in SoV as “Brand B is a premium serum.” The dashboard celebrates the mention. The brand’s positioning is being overwritten in real time.
2. Hallucinated claims pass undetected. Sometimes AI assistants invent facts — a wrong ingredient, a wrong country of origin, a wrong founder name. SoV scores the mention. NA flags the fabrication immediately by failing the Correctness test.
3. Tone collapse goes uncaught. A brand whose positioning is “scientifically rigorous” can be described by AI as “good for beginners.” The phrase isn’t wrong, exactly, but it’s off-brand. SoV cannot detect tone drift. NA, with a full ground-truth positioning document to compare against, can.
4. Coverage decay is invisible. A brand whose AI mentions used to come 80% from owned content can drift to 30% over six months as third-party content multiplies. SoV doesn’t track Coverage at all. The brand has lost authorial control over its own description and the dashboard reports nothing.
5. Cross-platform divergence is hidden. A brand can have NA 0.85 on ChatGPT (which has indexed the brand’s recent positioning) and NA 0.40 on Gemini (which is using older retail data). SoV averages them or reports them as platform-level mention counts; the per-platform NA gap, which is the actual investigation lead, is invisible.
How to start measuring Narrative Accuracy
You don’t need to wait for a vendor to ship a polished NA dashboard to start measuring. The minimal workflow is doable manually for one brand-week, automatable thereafter.
Step 1: Build the ground-truth positioning document. A short, durable artifact (one page, ten points) covering: positioning, target persona, price tier, ingredient claims, country of origin, brand voice descriptors, what the brand explicitly is not. This is what NA scores against. Brands that don’t have this codified will fail the Correctness test for reasons unrelated to AI — they don’t actually agree internally on what they are.
Step 2: Run twenty real shopper queries across four AI assistants. Not category keywords; specific shopper-voice queries. “What is Brand B?” “Is Brand B worth buying?” “How does Brand B compare to Glossier?” “What’s in Brand B’s serum?” Capture full text responses. Tag with timestamp, platform, query.
Step 3: Extract every claim the AI makes about the brand. A “claim” is any factual or descriptive statement: ingredient, price, target persona, positioning, comparison, origin, tone descriptor. Strip out non-claim filler (transitional phrasing, the AI’s hedging language).
Step 4: Score each claim on Correctness (0/0.5/1). 0 = wrong; 0.5 = directionally right but off; 1 = matches ground-truth. Correctness at the brand level is the weighted average across all claims, weighted by claim frequency.
Step 5: Trace each claim to its likely source (Coverage). This is harder; the AI doesn’t tell you where its claim came from. Heuristics: does the same wording appear in the brand’s product page schema? In a verified review snippet? In an old retail listing? In a competitor’s content? Best-fit attribution; a 0/1 binary at the claim level. Coverage at the brand level is the share of claims traceable to owned content.
Step 6: NA = Coverage × Correctness. Track over time, by platform, by query type.
The first two manual cycles will be revealing. Most brands discover that their NA is significantly lower than their SoV would suggest, and that the gap is widest precisely where they care most — premium positioning, ingredient claims, target persona.
Why this metric matters now, not later
There are three reasons NA cannot wait until SoV runs its course.
The training-data feedback loop is one-way. Every quarter spent letting inaccurate descriptions accumulate makes the next quarter harder. Coverage that reaches the AI from owned content compounds; coverage from third-party error also compounds. The brand that starts measuring and intervening on NA in 2026 is in a different position by 2027 than the brand that waits for SoV to obviously break.
The delegated mode raises the cost of bad NA every quarter. As more shoppers hand their buying jobs to personal AI agents, the AI’s description of your brand is no longer a perception layer the human can correct for — it is the operating input the agent acts on. Sloppy NA in 2025 was brand erosion. Sloppy NA in 2027 is invisibility to the half of the channel that has gone delegated.
Premium positioning is what AI most misrepresents. AI assistants tend toward median description; subtle premium cues (“considered,” “minimalist,” “clinical-grade”) get flattened into generic ones (“popular,” “well-regarded,” “affordable”). Premium brands are the ones whose positioning is the most defensible business asset and the most fragile in this channel. NA is the lever to defend it.
The dashboard you measure shapes the brand-team work. A team that reports SoV will optimise for mentions. A team that reports NA will optimise for the brand showing up correctly — schema-rich product pages, verified reviews, owned content that the AI can quote, ground-truth documents that the AI can cite. That work is what builds Coverage. Coverage builds Correctness. Correctness builds brand value. SoV doesn’t ask for any of it.
The measurement triangle
Yesterday we wrote about Purchase Likelihood — the revenue-layer metric. Today, Narrative Accuracy — the brand-value-layer metric. There is a third metric the AI Purchase Channel demands: Intervention Impact — the action-effectiveness metric. Did the content you published, the schema you added, the review you surfaced, actually move PL or NA? II is what closes the loop between work and outcome.
Together — PL, NA, II — these three metrics describe what the AI Purchase Channel needs and what visibility, ranking, share of voice, and other inherited dashboards do not. PL maps to revenue. NA maps to brand value. II maps to operating effectiveness. Each measures a layer the others can’t see. Brand teams running the channel seriously will, over the next twelve months, end up reporting all three.
The instrument any brand team measures with becomes the instrument they optimise for. The dashboards selling SoV and visibility are measuring the wrong thing on both axes. Switch the metric and the work follows.
What this means for the brand team this week
A premium brand reading this essay should, by end of week:
- Locate the ground-truth positioning document — or write it if it doesn’t exist.
- Run the twenty-query manual NA audit across ChatGPT, Gemini, Perplexity, and Google AI Overviews. Score Correctness. Estimate Coverage. Compute NA.
- Compare the NA number to the SoV number their current dashboard reports. The gap is the question.
- Identify the highest-frequency inaccurate claim and trace its likely source. That source is the first intervention target.
The brand team that does this in May 2026 will be ahead of every brand team that waits for vendors to ship NA dashboards in Q3 or Q4. The cost is one focused afternoon. The return is twelve months of compounding brand-equity protection in the channel that’s eating the next decade of consumer purchasing.
Frequently asked
What is the AI Purchase Channel? The AI Purchase Channel is the new buying channel where the cognitive work of buying — research, comparison, defensibility checks, choice — has moved out of the shopper’s head and into AI. Sometimes the shopper stays in the conversation with an AI assistant (the assisted mode); sometimes they hand the whole job to a personal AI agent (the delegated mode). Either way, the decision happens inside AI. (Read the Field Guide.)
What is Narrative Accuracy? Narrative Accuracy is the degree to which an AI assistant’s claims about your brand — its ingredients, positioning, price point, persona, provenance — are true to your owned positioning. It decomposes into Coverage (share of AI claims traceable to your owned content) and Correctness (share of AI claims that match ground-truth positioning). NA = Coverage × Correctness.
Why does NA matter more in the delegated mode? In the assisted mode, the AI’s description of your brand is read by a human who can correct for sloppy framing with their own taste. In the delegated mode, a personal AI agent reads the AI’s description and acts on it directly — there is no second chance. A 0.5-Correctness claim is forgivable for a human reader and disqualifying for an agent whose brief is precise. As the delegated share of the channel grows, sloppy NA stops being brand erosion and starts being invisibility.
What’s wrong with measuring Share of Voice in the AI Purchase Channel? SoV inherits a PR-era assumption that mentions are roughly fungible. AI-generated mentions are not — they can be accurate, neutral, inaccurate, or fabricated, and SoV counts all four identically. A brand can have 80% SoV and 20% NA: heavily mentioned, badly described. SoV doesn’t see the brand-equity erosion happening underneath the mention count.
Why is the training-data feedback loop dangerous? AI-generated descriptions get cached, shared, screenshotted, indexed by search engines, and eventually fold back into training corpora. Every inaccurate description today becomes a higher-base-rate description tomorrow. Brand drift in this channel compounds, and the brand’s marketing budget cannot outpace the loop unless the brand actively measures and intervenes on NA.
How does Narrative Accuracy relate to Purchase Likelihood? Purchase Likelihood is the revenue-layer metric — does the AI recommend you. Narrative Accuracy is the brand-value-layer metric — when the AI talks about you, is what it says true. A brand can have high PL and low NA (recommended often, described wrong) — and that combination erodes the brand-equity asset that justified the recommendation in the first place. Both metrics are needed.
Is there a third metric we should track? Yes — Intervention Impact (II), the action-effectiveness metric. After publishing content, adding schema, surfacing reviews — did PL or NA actually move? II closes the loop between work and outcome. Together, PL + NA + II are the measurement triangle the AI Purchase Channel demands. We’ll cover II in a future essay.
Read the AI Purchase Channel Field Guide — the canonical category essay.
Slingso is the AI Purchase Channel Manager — the AI team member your brand onboards to monitor, defend, and grow Narrative Accuracy continuously. Monitor → Analyse → Create → Approve → Measure, on a loop.