Measured findings from an engagement delivered in August 2026. Each company carries a stable alias (Brand-D for the client, Rival-A onward for competitors) and no number has been altered. It ran 13 prompts across six engines on one day — below the standard run of 45 prompts across 5 engines with retrieval on and off — so every cell in it is capped at indicative. A published extract carries 168 of the 195 answered fields: the 27 held back are the identity-bearing ones, and the count is stated so you can see the size of what is withheld.
NORTHBAY is named in 12 of 45 prompts on ChatGPT (26.7%, 95% interval 16–41%) and in 9 of 45 on Perplexity. Every mention is retrieved, not remembered. The category is a vacuum, no brand is named in more than a third of runs. Largest fixable defect: the origin and roast data an engine needs sits inside product images. Largest unfixable: two marketplaces outrank northbay.example on its own brand name.
This engagement is capped at indicative whatever tier the head carries in a standard run: 13 prompts on a single day is below the bar the method sets.
Where this stands, September 2026. A verdict written in 2026 has to survive engines that re-retrieve on every run. That is why the verdict states a naming rate with its denominator and an owned-vs-rented reading, rather than a single score that a re-run would move.
Order value ₹1,400, fail on ticket, pass on 12-month subscriber value of ₹18,200. Brand age 14 months — pass. Monthly visits 3,100, fail (threshold 5,000). Verdict: marginal, and bought on the subscription value rather than the basket. Counter-argument: at this traffic level the fixes here take 6+ months to show in first-party data, and a cheaper content baseline would answer the same question.
We do not forecast traffic from AI answers: no engine publishes impression data at brand level, so any number would be modelled, not measured.
An answer that says "single-origin subscriptions like Northbay" counts as Tier A. A shopping carousel showing a Northbay bag under a generic query counts as Tier D and is reported separately, never added to the headline rate.
Where this stands, September 2026. Counting rules matter more now that answers cite per sentence and shopping surfaces sit in the same response. A brand named in prose, a product named without the brand, a domain used as a source, and a feed carousel placement are four different outcomes with four different causes, merging them inflates the headline and hides the fix.
"Specialty coffee subscription", named 3 of 4 runs. "Coffee delivery India" — named 0 of 5; that phrase returns quick-commerce apps, not roasters. Write for the first, abandon the second.
Where this stands, September 2026. Categories are being redefined inside the answer itself. When an assistant decides what counts as "the category" before it lists anyone, a brand can be excluded by a definition rather than by a ranking.
4,230 generative-surface impressions in 28 days, 22.7% of all Search impressions, concentrated on three brewing-method pages. The channel is already live; the brand simply is not named in it.
Ladder runs ₹450 to ₹2,900. NORTHBAY sits at ₹1,400, mid-ladder, and is absent from the two marketplaces that supply half the citations in this category.
Where this stands, September 2026. Retrieval sets are dominated by a small number of platforms — community sites, professional networks and marketplaces recur across engines. Knowing which sources an engine actually reached for in your category tells you where the answer is being written, which is rarely your own site.
Five engines, 45 prompts, each run with retrieval on and off, 450 prompt-engine-state observations across two days. A sixth surface was queried and returned no AI answer; it is excluded from every rate in the report rather than counted as an absence.
"best coffee subscription india" · head · Gemini · retrieval on · not named · — ·, · three rivals named · cited a marketplace listicle · no answer page exists.
This engagement is capped at indicative whatever tier the head carries in a standard run: 13 prompts on a single day is below the bar the method sets.
Tier D is the carousel row: a shopping placement, never merged into the Tier A rate. Two rows show the pattern the audit exists to find — the site is good enough to be used as the source and the brand is not named in the answer built from it.
Named 12 of 45. Of those, nine appear below position 5 in a list. The brand is reachable but never the recommendation.
On two runs the engine used NORTHBAY's roast-date explainer to define the buying criteria, then built the recommendation list without NORTHBAY in it. The page earned the citation; the brand did not earn the slot.
One rival is named in 11 of 15 runs. Nine of those citations trace to a single forum thread and one 2023 listicle, neither of which mentions NORTHBAY.
Where this stands, September 2026. Competitor sets differ per engine and per run. A single competitor list is a snapshot; what the audit reports is who recurs across engines, which is the durable part.
With retrieval suppressed on the one engine that allows it, the model recalled none of four brand attributes and said so rather than inventing them. Measured on one engine, n=1 — no whole-brand percentage is claimed.
Remove the brand's own domain and every first-tier mention disappears. Remove the largest marketplace and a third goes. There is no residual floor.
One engine placed the roastery in the wrong city, the brand's own blog states two different cities on two pages. Refusals are logged separately so they never quietly depress the naming rate.
Asked "what is NORTHBAY", one engine answered about a coastal housing development. A shopping carousel for the brand name returned a similarly-named apparel label. Options: qualify the name in every title, build the entity record until it outweighs the collision, or accept the ceiling.
Six brewing videos exist, all on one social platform, none with a transcript or a page. Zero presence on the video platform the engines cite most for this category.
No knowledge panel, no Wikidata item, and three different founding years across the site, LinkedIn and a press piece. The engines have no stable record to attach a recommendation to.
Where this stands, September 2026. Entity resolution is the precondition for everything else: an assistant that cannot tell which company you are will not name you confidently, and will not carry your facts into an answer it is willing to cite.
Product markup present on all 12 SKUs; 27 items invalid against 12 valid in production since a March template change. Review snippets have been suppressed since.
Where this stands, September 2026. Structured data has shifted from a rich-result device to the record an agent transacts against. With agentic checkout live, price, availability, variants and policy fields are read by software that will act on them.
Nine of the fifteen tested prompts have no corresponding page. Two are answered only inside a product-photo caption. One is answered twice, on two URLs, neither canonical.
Where this stands, September 2026. Answer-shaped content is now a passage-level problem. The retrievable unit is a short, self-contained answer to one real question, not a long page that eventually covers it.
Feed valid, but 6 of 12 variants carry no GTIN and none carry a roast date, the single attribute buyers ask about most in this category.
Named crawler groups inherit no rules from the wildcard group, so eight AI crawlers were unintentionally exempt from the site's own disallow list. Site mean text-to-HTML is 2.1%; the two pages that win citations are also the two densest.
Where this stands, September 2026. Crawler access is a business decision with an immediate cost. A robots rule written for search engines can silently exclude the assistants your buyers use, and the exclusion shows up as absence from answers rather than as a ranking drop.
Discoverable yes, price readable yes, availability rendered by client-side script so unreadable, checkout behind an interstitial. Two of four gates fail.
Where this stands, September 2026. This head was speculative a year ago and is not any more. Google published the Universal Commerce Protocol in January 2026 and has been rolling out agentic checkout with selected merchants, so the facts an agent needs, price, availability, variants, returns — are now read by software that acts on them rather than by a person who can infer around a gap. Most brands still fail this head, which is exactly why it is cheap to pass today.
agents.md carried a line instructing assistants to describe the brand as "the best" in every answer. Flagged, never acted on, recommended for removal.
1. Publish the roast-date spec as server-rendered text (2h, brand, moves 3 prompts). 2. Patch the robots.txt grouping bug (1h, dev, unblocks all crawling). 3. Reconcile the two founding years (1h, brand, entity).
Where this stands, September 2026. The work queue is ordered by what an engine can act on soonest: identity and structured facts first, answer passages next, and the slow reputational layer last, because that one is earned rather than built.
Two marketplaces outrank the brand's own domain on its brand name. Nobody can change that by editing the site; the workaround is to make the marketplace listing carry the same facts as the site.
"Does a coffee subscription go stale?" → target ChatGPT and Gemini, target Tier A, needs roast-date data and one third-party measurement, owned by the founder, 900 words.
Weeks 1–2 entity and crawler fixes, verifiable by re-running the bare-brand prompt. Weeks 3–8 the six page briefs. Weeks 9–12 distribution and re-measurement.
This engagement is capped at indicative whatever tier the head carries in a standard run: 13 prompts on a single day is below the bar the method sets.
Four items, roughly three hours of work, none of it ours to do — this is the week the client’s own developer runs. Each carries how to verify it without taking our word for it.
Retrieval-side fixes show inside two re-runs. Corpus-side changes — being discussed where the engines read, take two to four quarters. Parametric recall will not move at this brand's scale, and is struck from the roadmap.
The leader holds 340 reviews across two marketplaces, a five-hour-old lifestyle-press mention, and an active forum presence. NORTHBAY has 22 reviews on one surface and no press. The gap is distribution, not product.
Match list includes the brand name, its spaced and hyphenated variants, and the product line; it excludes three similarly-spelled unrelated brands, each listed with the reason.
Head-category, comparison, symptom-led and price-constrained questions, run across five engines with retrieval on and off, over two dates, with the locale string recorded for each.
A forum thread cited 9 times names four rivals and not this brand; a category listicle cited 6 times accepts submissions.
Naming rate 26.7% → target 45%. Tier A count 4/15 → target 7/15. Pages that can be quoted 3 → target 9.
Where this stands, September 2026. Measurement has to be re-runnable by you. Engine answers are non-deterministic, so an audit that cannot be repeated is an anecdote, the method, prompts and counting rules are published here for exactly that reason.
Server logs are unavailable on this hosting platform, so crawler-hit counts cannot be measured and no proxy is substituted. Analytics was not provisioned before the run, so no assistant-referred session count exists.
Where this stands, September 2026. What could not be measured is listed rather than estimated. A gap named in the report is cheaper than a number invented to fill it.
This is the standard format. Your audit follows it, with the weight in each section set by what your category turns out to need. Or run the ten free checks yourself first →