Korpora measurement · Eight Sleep · sleep technology
Eight Sleep runs the most ads in its category.
Ask AI cold, and it barely surfaces.
932 Meta ads, the most in its cohort. Unprompted AI recall falls from 96.3% named to 18.5% on open discovery.
Named in a comparison, AI surfaces Eight Sleep 96.3% of the time. Strip the brand from the question and unprompted recall falls to 48.1%, and to just 18.5% on a pure open-ended discovery query. It runs 932 Meta ads, the most in its cohort, but that attention lives on channels the models never read. Closing the gap means earning presence on the surfaces they do read.
932 Meta ads, the most in its cohort. Unprompted AI recall falls from 96.3% named to 18.5% on open discovery.
Eight Sleep · public dashboard
Eight Sleep's AI mindshare, live
3 headline views
v3 snapshot · 2026-05-29
Where Eight Sleep stands in the answers AI assistants give buyers, measured against its sleep-tech cohort. Tab through the findings; each number is real, pulled from the v3 run on 2026-05-29, and re-pulled weekly on a paid engagement.
Action cards
What to ship before the next training corpus closes.
Nine cards, ordered by qualitative lift potential (discovery-gap evidence weight × ship feasibility × time-to-corpus). Each carries a paste-ready engineer prompt; the discovery buyer-guide card also carries a brand-voice growth prompt. The battery cutoff was 2025-12-01, so content shipped now is building toward the cohort after next, roughly 6–12 months out.
Active
9 cards · 0 shipped
In-flight and queued. Every card is an open recommendation today; on a paid engagement they push into the team's tools and ship against this board.
Shipped
0 cards executed
Public build history of the Eight Sleep engagement. Empty because this is a prospective measurement and nothing has shipped yet. The slot is wired, not decorative.
No cards executed yet
The first executed card appears here the minute it pings back through the protocol above.
How we work
Two steps. You are on step one.
Korpora does not run these for everyone. Eight Sleep cleared the internal fit review (scale, vertical, AI-channel relevance, execution capacity), so Korpora ran the measurement and shipped this dashboard. There was nothing to request and nothing to buy. What happens next comes down to one thing: whether the measurement in front of you earns the engagement.
- 1
Prospective measurement (you are here)
This is it. Korpora ran the full cross-channel measurement on Eight Sleep and shipped the result as the dashboard you are reading: cross-stream findings, ranked action cards (each with a paste-ready engineer prompt and a brand-risk callout), full methodology disclosure, and the Ask Korpora chat for follow-up questions. Run once, on us. Nothing was requested and nothing is owed. You read it and decide whether the measurement earns the engagement. See /ssense and /workwhile for the same treatment on other brands.
- 2
Paid engagement (what continuing looks like)
When the measurement earns it, continuing is a paid engagement scoped to Eight Sleep's cohort and depth: weekly re-measurement across every stream and every model, Western and Chinese, fresh action cards as the surfaces move, an alert the moment a rival accelerates into the next training corpus, and an operator who can run the moves alongside your team. There is no fixed self-serve tier; every cohort prices differently. Apply with a work email and we come back with scope and pricing, no call required to get the number.
Ask Korpora
Tell Korpora what you need
Methodology
How we measure, in full.
Every number on this page traces back to one of the streams below. Expand any section for the method, the headline-metric definitions, the live streams themselves, and the v3 caveats.
1 · How we measure
How each number on this page is produced. Every stream re-runs weekly on a paid engagement:
- AI assistant recall: buyer-decision queries (subject-blind and subject-named) run through three frontier models, 108 cells this cycle, Wilson 95% confidence range on every share. Organic (unprompted) reported separately from evaluation (named).
- LLM citation graph: which domains Google AI and ChatGPT cite for sleep-tech buyer queries, and how often each brand's own domain is cited. From DataForSEO LLM Mentions.
- Reddit corpus velocity: earned brand mentions across five neutral sleep and biohacking subreddits, the third-party surface AI cites most for this category, with recent publication velocity measured against a matched pre-cutoff window. Posts via the Arctic Shift archive.
- YouTube corpus presence: brand presence across the live sleep-tech video corpus, credited where a brand is named in a video's title, description, or transcript and deduped to earned uploads. youtube.com is the domain Google AI cites most for these buyer queries, so it feeds model memory the way Reddit does. Via the live YouTube pull.
- Creative landscape: competitive Meta ad inventory (Facebook + Instagram). Active ad volume per brand, from the live GetHookd ad library.
- Instagram organic: caption + engagement signal per post, content format mix. Via Apify (live), shown in the attention-footprint stream.
- TikTok organic: engagement-per-video, format pattern, the fastest-growing upstream for next-cycle editorial coverage. Via Apify (live), shown in the attention-footprint stream.
- Demand axis: branded AI search volume per brand, the demand-side counterpart to the supply-side mindshare battery. From DataForSEO AI search volume.
2 · How to read the headline metrics
- Organic owned share (48.1% [Wilson 95% CI 35.4–61.1], the metric that drives action)
- Eight Sleep was surfaced by AI without being named in the prompt, in 26 of 54 subject-blind cells (open-ended discovery + use-case questions). This is unprompted recall: a leading proxy for AI-driven discovery, not a measured conversion. That 48.1% is a blend of two unprompted framings: 77.8% (21/27) when the question anchors on a use-case, and 18.5% (5/27, Wilson 95% CI 8.2–36.7%) on pure open-ended discovery. Same “no brand named” rule, two very different framings: the discovery half is the cold-start liability, the use-case half is already healthy. The CI is wide (±13pp) because the subject-blind sample is 54 cells; a 5× larger battery tightens it materially.
- 50.0pp organic↔evaluation spread = cold-start classification
- The gap between organic (48.1%) and evaluation (98.1%) is 50.0 percentage points. A spread >40pp is what Korpora flags as cold-start (our own working threshold): the model engages with Eight Sleep when told to, but has limited independent signal to reach for it unprompted. Treat the 98.1% evaluation number as mostly query-framing artifact, not real buyer signal.
3 · The live measurement streams
Five data streams plus a demand axis, re-run weekly on a paid engagement. Streams 3 and 4 (Reddit, YouTube) are the corpus-feeding surfaces models actually read; Stream 5 collapses the three attention channels (Meta ads, Instagram, TikTok), none of which feed the training corpus, into one panel that sizes the attention-vs-memory gap.
Stream 1 of 5 · live data · Korpora 108-cell battery · 2026-05-29
AI assistant recall composition
Buyer-decision queries across 4 framings × 3 frontier models (Claude Sonnet, Claude Haiku, GPT-5.5) × 3 rounds, 108 cells total. Subject-blind cells (organic: discovery + use-case) reported separately from subject-named (evaluation: direct-comparison + differentiation).
Organic mindshare composition (54 subject-blind cells)
52% of subject-blind cells are either contested or unclaimed, the mindshare we would treat as addressable over the next two corpus cycles. The category is fragmented; no brand is dominant unprompted.
Organic owned share 48.1% [Wilson 95% CI 35.4–61.1]. How to read this and the 50.0pp cold-start spread is in the method notes below.
Subject share by query framing
The deficit concentrates in one place: discovery framing at 18.5%(5/27, Wilson 95% CI 8.2–36.7%). Eight Sleep is recalled in nearly all the named comparisons (direct-comparison 96.3%, differentiation 100%) and most use-case prompts (77.8%), but buyer journeys that begin with an open-ended question surface it less than one time in five. The interval is wide on a 27-cell slice, but it triangulates with two independent streams that point the same way: the demand-side branded-search deficit and the mid-pack earned-citation position. That is the single largest agent-layer liability and the target of Cards 1, 2, 3, 4, 6.
Subject share by model (aggregate, all framings)
13.9pp spread across models (sonnet 80.6%, gpt-5.5 72.2%, haiku 66.7%), inside the “aligned” threshold, so no single model is a meaningful outlier. Haiku is the floor; Card 5 (the sleep-stage methodology post) targets lifting lower-capability models toward parity.
Aggregate co-occurrence share · reference-only (within Eight Sleep's battery)
Read this as reference, not headline. Every bar here is co-occurrence inside Eight Sleep's own battery. A competitor's share is how often it surfaces in cells whose queries were built around Eight Sleep, not that rival's own mindshare. Sleep Number at 38.9% does not mean Sleep Number owns 38.9% of the category; it means Sleep Number appeared in 42 of these 108 cells. So do not read two competitor bars against each other as if they were each brand's standing. The subject bar is inflated on top of that, because subject-named cells are included. A rival's true number needs that rival run as the subject of its own battery (per-rival subject-own batteries, as we did for the SSENSE cohort, are a v4 step). Sleep Number is the most frequently surfaced rival; Tempur-Pedic never surfaced.
Cross-round stability: this reference aggregate holds at 73.1% ± 1.6pp across 3 rounds (75.0 / 72.2 / 72.2), stable to query selection. That stability is of the reference-only co-occurrence number, not the 48.1% headline. Source: eight-sleep-v3-2026-05-29 battery (36 queries × 3 models × 1 trial = 108 cells). Trials-per-cell is 1 this cycle; a paid engagement raises it for tighter within-cell CIs.
Stream 2 of 5 · live data · DataForSEO LLM Mentions · 2026-05-29
LLM citation graph
For the buyer query “best bed cooling system”, which domains do Google AI and ChatGPT cite as sources, and how often is each brand's own domain cited?
Google AI · top cited domains
ChatGPT · top cited domains
Own-domain citations (domain-scoped LLM Mentions)
| Brand | Domain | Google AI | ChatGPT |
|---|---|---|---|
| Oura | ouraring.com | 9,397 | 143 |
| Sleepme | sleep.me | 3,595 | 144 |
| Eight Sleep | www.eightsleep.com | 734 | 33 |
| BedJet | bedjet.com | 304 | 27 |
| Sleep Number | sleepnumber.com | 2 | null |
| Whoop | whoop.com | 1 | null |
| Tempur-Pedic | tempurpedic.com | null | null |
AI cites third-party authority (YouTube, Reddit, Sleep Foundation, Tom's Guide, Forbes, Healthline) for sleep-tech queries. www.eightsleep.com earns 734 citations on Google AI, 3rd of 7, but trails Oura (9,397, ~13×) and Sleepme (3,595, ~5×), and is thinner still on ChatGPT (33). For the brand running the most ads in the cohort, that is mid-pack earned presence on the surfaces AI reads. The fix is two-pronged: feed the third-party surfaces AI cites most (Cards 1, 3, 4, 6) and make www.eightsleep.com more extractable with structured FAQ data (Card 2).
Data confidence. Top-domain counts are query-scoped to one representative buyer query; the own-domain counts are aggregated-metrics scoped to each brand domain, matched on the exact host. Eight Sleep is scoped to its canonical host www.eightsleep.com; the apex eightsleep.com returns null, so an earlier apex-only pull understated it. The robust signal is the rank (Eight Sleep 3rd of 7, trailing Oura and Sleepme by 5–13×); single-query per-source ordering is not statistically reliable on its own.
Stream 3 of 5 · Reddit corpus velocity · paused
Corpus-feeding velocity (Reddit)
Stream 2 shows reddit.com is the domain AI cites most for sleep-tech buyer queries (#1 on ChatGPT, #2 on Google AI), so Reddit is a surface the models actually read, and that citation behaviour is measured there. Per-brand Reddit corpus ingestion is paused pending a commercial data license, so the footprint and velocity figures are held back here. This stream returns, with the full per-brand footprint and post-cutoff velocity, once the licensing is in place.
Stream 4 of 5 · live data · YouTube category corpus · 2026-05-29
Corpus-feeding presence (YouTube)
Stream 2 showed youtube.com is the domain Google AI cites most for sleep-cooling buyer queries: #1, at 761 citations, ahead of reddit.com. It is the other surface in this lineup models actually read, so presence in the YouTube category corpus compounds into model memory the way Reddit does and paid attention (Stream 5) does not. We pulled the live corpus, 36 videos across three buyer queries with 35 transcripts, and counted brand presence by video, credited where named in title, description, or transcript.
Against its real smart/climate cohort, Eight Sleep leads earned presence: named in 12 of 36 videos (33%), ahead of Tempur-Pedic (7), Sleep Number (6), Sleepme/ChiliPad (4) and BedJet (4); Oura is absent. This is the one surface in the dashboard where Eight Sleep is not behind its cohort. The caveat: the broader category corpus is owned by foam-mattress roundups (Saatva appears in 20 videos, Helix and Brooklyn Bedding 18 each), so the buyer searching "best mattress" meets a foam frame first.
Owned-vs-earned dedup · why raw view-share misleads
A naive view-weighted share would report Eight Sleep at 99.2% of the corpus (64.9M of the 65.4M-view pool) and call it total dominance.
It is one video. The Pod 5 launch ad, an owned upload, carries 64.0M views, 98.5% of every view Eight Sleep is credited with. Strip the owned ad and Eight Sleep's earned reach, the 11 third-party videos that actually feed the citation graph, is 949K views.
This is why Korpora separates owned from earned before it reports a number. The raw scrape says 99%; the honest earned figure is 949K. The dedup is the measurement, not the pull.
Reading the transcripts is what makes the count honest. Title and description alone name Eight Sleep in 8 videos; the 35 transcripts surface 4 more (to 12), a 50% lift, because the comparison that matters, the spoken "Eight Sleep vs ChiliPad" inside a review, lives in the audio, not the metadata. The corpus is also gatekept: the top five channels hold 42% of it, led by Mattress Clarity and Sleep Doctor (six videos each), with the Wall Street Journal, AuthenTech and Forbes the high-authority carriers.
Read against Stream 2 (youtube.com is the #1 domain Google AI cites for these queries), earned YouTube presence is corpus supply that compounds into AI memory. Eight Sleep leads its cohort here, but on thin earned reach (949K) carried by a handful of review channels, while its 64M owned ad feeds none of it. Deepening genuine presence in the roundups those channels publish is the supply-side lever the citation graph reads. An inference from two aligned signals, not a proven mechanism.
Not a placement recommendation. This stream measures earned presence; the legitimate lever is being genuinely worth a review channel's roundup, not buying the slot.
Data confidence. Apify YouTube scraper, 2026-05-29: top ~12 results by relevance for each of three buyer queries ("best smart mattress", "best cooling mattress", "best mattress for hot sleepers"), 36 unique videos, 35 transcripts captured. Presence is credited where the brand is named in title, description, or transcript. Owned = Eight Sleep's brand channel or a paid placement; earned = third-party uploads. The view pool is dominated by one owned ad (see dedup), so video-count presence, not view-share, is the comparable metric. A single-pull snapshot, not a tracked series.
Stream 5 of 5 · attention footprint vs AI memory
Meta ads, Instagram, and TikTok are channels models do not read. None feed the training corpus, so none move organic AI mindshare directly. They are measured together here to size one gap: the attention Eight Sleep buys and earns on human channels, against how little of it compounds into the model memory measured in Streams 1 and 2. Read this as the attention-vs-earned-memory contrast, not as a mindshare input.
Paid signal · Meta ad inventory · GetHookd · 2026-05-29
Active Meta ad inventory
Count of unique active Meta ads (Facebook + Instagram placements combined) per brand, snapshotted from the GetHookd ad library.
Eight Sleep runs the most active ads in the cohort (932), ahead of Whoop (729) and Oura (567). Read this against Stream 2: the brand with the heaviest paid presence sits only 3rd of 7 on earned domain-level AI citation, out-cited 5–13× by Oura and Sleepme, which run fewer ads. Paid volume runs well ahead of earned AI presence here.
Data confidence. Single-timestamp snapshot, no trend curve yet. Active-ad counts combine Facebook + Instagram placements (not deduplicated). Weekly cadence on a paid engagement would add the trend line; v3 carries the single snapshot.
Organic signal · Instagram · captured 2026-05-29
Instagram organic engagement-per-post
Mean of (likes + comments) on the most recent 30 posts per brand. Via Apify. Handles verified against the returned account owner.
Eight Sleep ranks 6th of 7 (343/post), ahead only of Sleepme / ChiliPad (121). Whoop's 81,306 is 237× Eight Sleep, and even the legacy mattress brands out-engage it (Sleep Number 7.8×, Tempur-Pedic 4.6×). The shape is an engaged-but-small audience: Eight Sleep posts a higher comment-to-like ratio (34 comments on 309 likes) than any rival except Sleepme, but absolute reach is low. Same direction as the cold-start finding in Stream 1.
Data confidence. Most-recent 30 posts per brand. Bootstrap ±15% on the per-brand means; a single viral post moves the mean. Instagram hides exact like counts on some Eight Sleep posts, so the subject mean is taken over available counts only.
Organic signal · TikTok · captured 2026-05-29
TikTok organic engagement-per-video
Mean of (likes + comments + shares + saves) on the most recent 30 videos per brand. Via Apify.
Eight Sleep ranks 6th of 7 by engagement-per-video (29), ahead only of the dormant Tempur-Pedic account. It holds the 3rd-largest following in the cohort (53,500) but the lowest ER of any active account (0.06%): a large but inert audience. BedJet (18k followers) and Sleepme / ChiliPad (1.3k) generate 350× and 20× Eight Sleep's per-video engagement off far smaller bases.
Data gap: Tempur-Pedic. Its official TikTok (@tempurpedic_usa) is effectively dormant, 1 video and near-zero engagement, so it is excluded from the ranking rather than shown as a zero bar.
Data confidence. Most-recent 30 videos per brand (14 for Sleep Number, which posts infrequently). Engagement is volatile post-to-post; bootstrap ±20–40% on the per-brand means. Read ER% with care: it divides per-video engagement by follower count, so small accounts whose videos reach beyond their followers (ChiliPad 1.3k, BedJet 18k) post inflated ER. Engagement-per-video is the more comparable metric, and even it is reach-confounded. Account-age / cadence confound applies as in Stream 1.
Demand axis · live data · DataForSEO AI search volume · 2026-05-29
Branded AI search volume
Monthly AI search volume on each brand's single highest-volume brand term. The demand-side counterpart to the supply-side mindshare battery: how often buyers ask AI about each brand by name.
Eight Sleep's branded AI search demand (5,564/mo) trails Oura (6.9×), Whoop (4.9×), Sleep Number (2.9×) and Tempur-Pedic (2.8×). The brand is not yet a top-of-mind name buyers ask AI about by default. Demand and supply (Stream 1's 18.5% discovery share) point at the same open-question gap.
Method note. We show each brand's single highest-volume brand term, not a sum of variants. Eight Sleep's “eight sleep” variant adds another 1,000/mo; summing variants does not change the rank order. Sleepme's strongest term is its ChiliPad sub-brand (419); the “sleepme” term alone is 39.
4 · v3 notes & caveats
- Query battery. v3 ran buyer-decision queries inferred from the category, not Eight Sleep's real on-site search terms or demand-side analytics. A paid engagement swaps the inferred queries for the terms buyers actually use, so the organic-mindshare number and the per-query gaps get materially more representative.
- Aggregate is reference-only. The 73.1% aggregate competitor share mixes subject-blind and subject-named cells and overstates the position; it is shown for back-comparison only. The headline organic number is 48.1% (subject-blind cells). Per-rival subject-own batteries (each competitor as subject of its own battery, as run for the SSENSE cohort) are a v4 step.
- Social handle verification. Instagram and TikTok organic are live. Each handle was verified against the account owner Apify returned, then corrected where the obvious slug was wrong: Sleepme's branded account is @chilipad (not @sleepme), BedJet is @bedjetsleep (the bare @bedjet is an unrelated person), and Tempur-Pedic's TikTok (@tempurpedic_usa) is dormant (1 video), so it is excluded from the TikTok ranking rather than shown as a zero.
- Card lift quantification. Cards give the mechanism, the hypothesis, and the re-measurement criterion, not yet a calibrated “N points over M corpus cycles” forecast. The ordering is a qualitative lift ranking.
Re-measurement cadence on engagement: weekly across all streams. The Korpora measurement model and the channel surfaces it measures both evolve continuously; weekly captures inflections that quarterly would miss.
Methodology question or factual correction? Ask in the chat above.
Related: full methodology · home