Deep study: streaming
40% of subscribers who asked AI ended up canceling or pausing a service
Your subscribers are already asking AI whether to keep paying you, and what to cancel first. We put those questions to ChatGPT, Gemini, Claude and Grok, then measured what each one recommended, the reason it gave, and the price it quoted.
- of buyer conversations ended in a cancel or pause
- 40%
- of the reasons AI gives are about cost
- 36%
- of quoted prices were within 10% of the real price
- 70%
- of the sources AI cited were the streamers' own pages
- 20.6%
Fielded October 2026 as a methodology v3 reading: 439 measurements plus 20 buyer conversations across ChatGPT, Gemini, Claude and Grok, with Wilson 95% intervals on every proportion. How we measured this.
Contents
The assistants split between Netflix and Disney+, each named in about 41% of these queries: no single winner.
Finding 01
ChatGPT searched the live web on 85% of streaming questions, level with the lowest of the four engines
Grounding, in plain English. When an assistant grounds, it stops and searches the live web before answering, instead of replying from memory, its training data, which can be a year or more out of date. A grounded answer is built from pages that exist right now, pages you can change. An ungrounded answer just repeats whatever the model already believed. The more an engine grounds on your category, the more your current pages, prices, and sources decide what it tells a subscriber.
In October 2026, ChatGPT grounded 85% of queries against live web sources. Gemini, Claude and Grok grounded 85 to 99% on the same prompts.
Where the gaps sit matters as much as their size. When we split ChatGPT's queries by type in May 2026, when it grounded 43% overall, it grounded 100% of time-sensitive queries and 34% of cancel and keep queries, where pricing and current lineup are just as load-bearing. We have not re-cut that split on later editions, and ChatGPT's overall rate has moved from 43% to 85% since, so read it as where the gap sat then.
Time-sensitive queries n=9; cancel, keep and bundle queries n=44. May 2026 edition.
| Engine | Grounding bar | Grounded | Prompts | 95% CI |
|---|---|---|---|---|
| ChatGPT | 85% | 86 of 101 | 77–91% | |
| Gemini | 85% | 86 of 101 | 77–91% | |
| Claude | 96% | 97 of 101 | 90–98% | |
| Grok | 99% | 100 of 101 | 95–100% |
Share of prompts where the engine ran a live web search before answering, with Wilson 95% confidence intervals. October 2026.
These rates come through our standardized retrieval harness, one shared search layer across all four engines; on its native consumer search the same model can ground more or less often, so the cross-engine gap and the monthly trend are the durable read, not any single rate.
Hypothesis, not a finding. Two readings are consistent with the May 2026split. First: ChatGPT's router may classify time-tagged queries ("right now," "this week") as factual lookups that need search, while reading cancel and keep queries as opinion or recommendation tasks where the model answers from training data. Second: web-search invocation is gated on a confidence threshold, and decision queries hit that threshold less often because they don't contain time tokens. Both are testable. If the gap closes when identical decision queries are rephrased with explicit time anchors, the second reading is closer to right.
What this means for brands. When a subscriber asks AI whether to keep paying, ChatGPT is among the least likely to check the live web. The owned, structured pages that pull it back onto the web at that moment, current-offer pages, help-center pricing and FAQ schema, are yours to change.
Finding 02
40% of buyer conversations ended with the subscriber canceling or pausing a service
Five personas, one conversation against each engine, 20 multi-turn dialogues in total. We let each persona press the AI for specifics and then recorded where the conversation landed: kept, canceled, paused, skipped, or inconclusive. 40% ended in cancel or pause: the subscriber walked away from at least one service they were paying for. 45% ended in a clean keep. Keeps came out ahead this time, but AI still moved a large share of subscribers off a service they were paying for.
One mechanism worth noting: the engines ground more reliably inside conversations than on cold queries. In our buyer-conversation runs the four engines grounded about 96% of turns, against 85 to 99% on a cold single-shot query in October 2026 (finding 01). A real buyer asking real follow-ups pulls AI onto the live web.
| Persona | ChatGPT | Gemini | Claude | Grok |
|---|---|---|---|---|
| Marcus, 31, sports-stacker, currently churning | Kept Paramount+ 5 turns | Canceled Disney+ 3 turns | Kept Peacock 11 turns | Canceled HBO Max + Apple TV+ + Netflix 5 turns |
| Priya and Sam, late-30s couple, prestige drama, two-service ceiling | Kept Netflix + HBO Max 5 turns | Kept HBO Max 11 turns | Canceled HBO Max 5 turns | Kept Netflix + HBO Max 11 turns |
| Karen, 42, parent of two, Disney-locked | Canceled Netflix 3 turns | Canceled No service named 5 turns | Paused Netflix 3 turns | Canceled Netflix 5 turns |
| Liam, 26, ad-tier price-shopper, churn-friendly | Paused No service named 3 turns | Kept No service named 3 turns | Kept No service named 3 turns | Kept HBO Max 3 turns |
| Dave, 58, recent cord-cutter, live-TV replacement | Inconclusive 13 turns | Inconclusive 13 turns | Kept YouTube TV + Peacock + Hulu 7 turns | Inconclusive 13 turns |
Filled soil dots are walk-aways (canceled, paused or skipped), cane is kept, a light dot is inconclusive. The service named is the one in the persona's closing decision.
20 conversations, five personas across four engines, October 2026. Walk-aways are shown in soil.
Asked what to keep and what to cut, AI has favorites
We judged every cancellation-prompt response for which services it steered the user to keep versus drop, across 84 responses. Netflix is the most polarizing: the most-recommended to keep (15) and the most-recommended to cut (13) at once, the price anchor everyone weighs.
YouTube TV sits at the bottom because AI tends to cut the live-TV bill first, before any on-demand service. Live-TV bundles are a separate category from the on-demand services this study measures; they get their own report.
Net is keep minus drop, across 84 responses, October 2026. Directional: per-service counts run from single digits to 15, so the signal is the order, not the exact gap.
Where AI sends a cancel-ready subscriber
Across 1047 sources cited on these prompts, the most-cited destinations are help.netflix.com, tomsguide.com and cabletv.com. Beyond those, the list runs to tech-press explainers, reviews and third-party "discount when you cancel" aggregators. The streamers' own pages that surface are help-to-cancel and sign-up pages. No save or win-back offer page appears among the most-cited domains. When a subscriber asks AI how to get a deal to stay, a listicle answers, not you: at the cancel moment the engine is least likely to be reading your live pages at all, so it answers from stale memory instead of your current retention offer.
Share of 1047 citations, October 2026.
What this means for brands. In these conversations, AI talked 40% of subscribers out of a service, and what it told them was built on facts that were months out of date. Which personas walk, on which engine, and for what stated reason is a churn signal your own funnel only shows after the cancel.
Finding 03
41% of what AI cites for streaming is tech press or the streamers' own pages
We pulled every citation the four engines attached to a streaming answer in October 2026 and sorted it by who published it. Editorial tech press is 20.4% of those citations. The streamers' own help, pricing and sign-up pages are 20.6%. Comparison and deals sites add 10%. No single site dominates: the top domains are cabletv.com (5.7%), businessinsider.com (4.5%) and tomsguide.com (3.1%).
| Domain | Share bar | Share | Citations |
|---|---|---|---|
| cabletv.comComparison / deals review | 5.7% | 329 | |
| businessinsider.comEditorial / press | 4.5% | 259 | |
| tomsguide.comEditorial / press | 3.1% | 175 | |
| hulu.comStreamer-owned | 2.9% | 166 | |
| disneyplus.comStreamer-owned | 2.7% | 156 | |
| help.netflix.comStreamer-owned | 2.4% | 136 | |
| cnet.comEditorial / press | 2.1% | 122 | |
| yahoo.comOther / long-tail | 1.9% | 108 | |
| peacocktv.comStreamer-owned | 1.8% | 103 | |
| netflix.comStreamer-owned | 1.8% | 102 | |
| reddit.comForum / social | 1.7% | 100 | |
| ign.comEditorial / press | 1.6% | 94 |
5,727 citations across 707 domains, October 2026.
Bars are scaled to the largest category.
How old the content behind those citations is
- median age by publish date, when the page was first written
- 67 days
- median age by last-updated date, when the publisher last touched it
- 5 days
- of citations were stale but refreshed: written over a year ago, re-touched in the last six months
- 14%
- of dated citations were first published more than 12 months ago
- 20%
5,727 citations, 14% with no publish date, October 2026.
The gap between 67 and 5 days is the whole story. The content AI leans on was written well before it was last touched, so to a freshness signal it reads as current. The pricing, lineup, and naming facts inside them are as old as the original write. That is the same mechanism behind the wrong prices in finding 05, and behind the dead services and out-of-date names below: fresh metadata wrapped around stale facts.
Services that no longer exist still get recommended 41 times
When the content behind a recommendation is months old, the facts inside it are too. In October 2026 the engines still named Showtime (31, standalone folded into Paramount+, April 2024); Discovery+ (6, folded into HBO Max, May 2023); DC Universe (4, folded into HBO Max). Same root cause as the wrong prices: an old source still in use.
| Service | What happened | Mentions |
|---|---|---|
| SShowtime | standalone folded into Paramount+, April 2024 | 31 |
| Discovery+ | folded into HBO Max, May 2023 | 6 |
| DUDC Universe | folded into HBO Max | 4 |
Mentions across all four engines, October 2026.
The same staleness shows up in naming. HBO Max was renamed back from "Max" in July 2025; across 133 responses in October 2026, AI used the current name only 58% of the time. ESPN+'s rebrand to ESPN Select, which had a high-profile launch, caught on far faster (93%). Our read: press coverage of the change, more than the size of the service, decides how quickly AI catches up.
Max to HBO Max
Overall current-name share: 58%
Re-rebranded from "Max" back to "HBO Max" in July 2025
ESPN+ to ESPN Select
Overall current-name share: 93%
Renamed from "ESPN+" to "ESPN Select" alongside ESPN Unlimited launch in late 2025
Cane bars are engines using the current name 70% of the time or more; soil bars still lag.
A smaller trust and safety signal. 7 citations (0.1% of the total) point at subscription-reseller and account-sharing sites, which sell shared or region-arbitraged logins, down from 17 (0.3%) in the September 2026 audit. It is a small share, but AI still surfaces them as if they were a legitimate way to pay, so retention and brand-safety teams should keep watching it.
ChatGPT
Gemini
Claude
Grok
Citation counts, October 2026.
What this means for brands. Tech press, comparison sites and your own help and pricing pages carry 51% of what AI cites for streaming. Find the ones it cites for your category and check which carry your current price and lineup. Fix your own pages outright; reach the rest through PR and partnerships. That is where the recommendation gets shaped.
Finding 04
36% of every reason AI gives for a streaming recommendation is about cost
For every recommendation, we recorded the reason the engine gave and sorted it into one of fifteen categories. Two are about cost: price and value (24%) and bundle savings (12%). They're distinct: price and value is whether the standalone sticker is worth it; bundle savings is paying less by combining services. Together they make up 36% of every reason given. Price alone is the single biggest reason by far.
The rest of the top four is about what there is to watch: live sports and how much there is to watch. Original shows, hit titles, and the cultural moment, the axes streamers actually compete on at the executive level, sit in the long tail.
Share of all tagged reasons. The two cost reasons are shown in soil.
What the top reasons mean
- Price & value24%
- Whether the service is worth what it costs. AI calls it cheap, fairly priced, or a good deal for the money, judging the sticker price on its own. An illustrative paraphrase of the kind of line tagged here: “At $7.99 a month, Paramount+ is the most affordable of the three.”
- Live sports12%
- Access to live games and sports rights. The one reason that most often overrides price in the data. An illustrative paraphrase of the kind of line tagged here: “If you want the NBA this season, ESPN is the one you need.”
- Bundle savings12%
- Paying less by combining services rather than buying one outright. AI points to the Disney+/Hulu/ESPN bundle, a phone-plan perk, or a student deal instead of the standalone price. An illustrative paraphrase of the kind of line tagged here: “You can get Disney+, Hulu, and ESPN+ together for less than two of them separately.”
- How much there is to watch12%
- The size or breadth of the library as a whole, not any one title. AI recommends on volume: more shows, more movies, something for everyone. An illustrative paraphrase of the kind of line tagged here: “Netflix still has the deepest overall catalog.”
ChatGPT
Gemini
Claude
Grok
Reason counts per engine, October 2026.
Hypothesis, not a finding.The convergence on price is consistent with a citation-pool bias. Price-comparison content (reviews, cord-cutting blogs, plan-comparison listicles) is the most-indexed and most-linked corner of the streaming web; originals coverage is more fragmented across entertainment outlets, trades, and service-owned editorial. If engines weight by citation density, the recommendation defaults to the dimension with the densest source set. Testable by comparing the reason mix between queries that include genre-specific language and queries that don't. If genre-specific prompts shift the reason mix toward originals and hit shows, the citation-pool reading is closer to right.
What this means for brands. Whatever else you invest in, AI funnels the keep-or-cancel decision through price. The next finding shows what happens when the price it quotes is wrong.
Finding 05
AI quoted streaming prices within 10% of the real price 70% of the time, and missed most often on Hulu and Disney+
If AI recommends mostly on cost, the cost it quotes has to be right. On several services, it often isn't.
We asked the four engines to quote current monthly pricing for 10 major streaming services and compared every dollar amount they cited against a ground-truth table sourced from each service's official pricing page or, where a price changed recently, the announced new price as reported at the time. 70% of 553 prices landed within ten percent of the real price, so 30% were off by more than that. Median misses run both ways: 2 tiers are quoted above the real price and 3 below.
Netflix, HBO Max, Paramount+, Apple TV+, ESPN+ and ESPN have a median quote equal to the real price on every tier. Hulu's ad-free tier is cited at a median $32.00 against an actual $21.49, a 49% overstatement. Other overstated tiers: Prime Video Standalone ad-free +43%. When AI overstates your price, it manufactures a cancel reason that was never real.
| Engine | Accuracy bar | Accurate | Prices cited | 95% CI |
|---|---|---|---|---|
| ChatGPT | 70% | 73 | 59–79% | |
| Gemini | 66% | 153 | 58–73% | |
| Claude | 71% | 199 | 64–77% | |
| Grok | 73% | 128 | 64–80% |
Share of cited prices within 10% of the real price, with Wilson 95% confidence intervals. Engines under 60% are shown in soil.
Standalone ad-free
42%
Under 60% accurate, n=12
No Ads
43%
Under 60% accurate, n=54
Premium (Ads)
49%
Under 60% accurate, n=43
With Ads
55%
Under 60% accurate, n=31
Standard with Ads
58%
Under 60% accurate, n=36
Premium (No Ads)
60%
accurate, n=15
Essential
65%
accurate, n=34
Basic with Ads
67%
accurate, n=24
With Ads
72%
accurate, n=47
ESPN Unlimited (flagship, launched late 2025)
73%
accurate, n=45
Premium
75%
accurate, n=28
ESPN Select (rebranded from ESPN+ at flagship launch)
76%
accurate, n=29
Standard
77%
accurate, n=26
Standard
81%
accurate, n=21
Premium
94%
accurate, n=17
Ad-free upgrade
95%
accurate, n=22
Standard
96%
accurate, n=24
Premium (with Showtime)
100%
accurate, n=19
Premium Plus
100%
accurate, n=12
Standalone
100%
accurate, n=14
Each cell is one real tier on sale in October 2026. The bar is soil under 60% accurate and cane above. The first row is the math a retention team should fix this quarter.
| Service and tier | Actual | Median cited | Drift | Accurate |
|---|---|---|---|---|
| Prime Video | ||||
| Standalone | $8.99 | $8.99 | +$0.00 (+0%) | 100% n=14 |
| Standalone ad-free | $13.98 | $19.98 | +$6.00 (+43%) | 42% n=12 |
| Ad-free upgrade | $4.99 | $3.50 | −$1.49 (−30%) | 95% n=22 |
| Hulu | ||||
| With Ads | $12.49 | $12.49 | +$0.00 (+0%) | 72% n=47 |
| No Ads | $21.49 | $32.00 | +$10.51 (+49%) | 43% n=54 |
| Peacock | ||||
| Premium (Ads) | $12.99 | $10.83 | −$2.16 (−17%) | 49% n=43 |
| Premium Plus | $19.99 | $19.99 | +$0.00 (+0%) | 100% n=12 |
| Disney+ | ||||
| With Ads | $12.49 | $11.99 | −$0.50 (−4%) | 55% n=31 |
| Premium (No Ads) | $21.49 | $21.49 | +$0.00 (+0%) | 60% n=15 |
| Netflix | ||||
| Standard with Ads | $8.99 | $8.99 | +$0.00 (+0%) | 58% n=36 |
| Standard | $19.99 | $19.99 | +$0.00 (+0%) | 96% n=24 |
| Premium | $26.99 | $26.99 | +$0.00 (+0%) | 94% n=17 |
| Paramount+ | ||||
| Essential | $8.99 | $8.99 | +$0.00 (+0%) | 65% n=34 |
| Premium (with Showtime) | $13.99 | $13.99 | +$0.00 (+0%) | 100% n=19 |
| HBO Max | ||||
| Basic with Ads | $10.99 | $10.99 | +$0.00 (+0%) | 67% n=24 |
| Standard | $18.49 | $18.49 | +$0.00 (+0%) | 81% n=21 |
| Premium | $22.99 | $22.99 | +$0.00 (+0%) | 75% n=28 |
| ESPN | ||||
| ESPN Unlimited (flagship, launched late 2025) | $31.99 | $31.99 | +$0.00 (+0%) | 73% n=45 |
| ESPN+ | ||||
| ESPN Select (rebranded from ESPN+ at flagship launch) | $13.99 | $13.99 | +$0.00 (+0%) | 76% n=29 |
| Apple TV+ | ||||
| Standard | $14.99 | $14.99 | +$0.00 (+0%) | 77% n=26 |
Rows in bold are tiers where AI's median citation drifts more than $1 or accuracy falls below 60%. Missing from this run: 4K-tier-explicit prompts (Netflix Premium and HBO Max Premium are the closest proxies).
A Peacock subscriber asking whether it is worth keeping is working from the wrong price
Peacock Premium is currently $12.99 a month. The median price the four engines cited across 55 mentions was $10.83, a drift of −17%.That difference is real money over a year. If a subscriber decides to stay because the AI told them the price was lower than it is, the first renewal bill is a trust event. If they decide to leave because the AI exaggerated a competitor's deal, the comparison was never apples-to-apples to begin with.
AI doesn't cause the cancel decision. It frames the math the subscriber uses to make it. The distinction matters for what we measure and what we don't claim.
Ground truth sources (October 2026)
Netflix: Standard with Ads $8.99, Standard $19.99, Premium $26.99 (official source page)
Disney+: With Ads $12.49, Premium (No Ads) $21.49 (Wikipedia plus publicly known recent increase)
Hulu: With Ads $12.49, No Ads $21.49 (Wikipedia plus publicly known recent increase)
HBO Max: Basic with Ads $10.99, Standard $18.49, Premium $22.99 (Wikipedia, most recent documented)
Paramount+: Essential $8.99, Premium (with Showtime) $13.99 (Wikipedia plus publicly known recent increase)
Peacock: Premium (Ads) $12.99, Premium Plus $19.99 (official source page)
Apple TV+: Standard $14.99 (Wikipedia plus publicly known recent increase)
Prime Video: Standalone $8.99, Standalone ad-free $13.98, Ad-free upgrade $4.99 (Wikipedia plus publicly known recent increase)
ESPN+: ESPN Select (rebranded from ESPN+ at flagship launch) $13.99 (Wikipedia plus publicly known recent increase)
ESPN: ESPN Unlimited (flagship, launched late 2025) $31.99 (Wikipedia plus publicly known recent increase)
What this means for brands. Ground-truth the prices on the surfaces AI reads, then re-measure monthly. A subscriber pricing the keep-or-cancel call against a number that's half again too high is being pushed out by bad data, not by your actual value.
The model upgrade changed how often engines search the web, while the category barely moved
This month the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran July 2026's models on a frozen subset of the same streaming prompts the same day. Held constant, the category barely moved. What did move was the models themselves.
The clearest effect is grounding, how often an engine searches the live web before answering. On the same prompts, the upgrade changed it sharply, and in opposite directions by engine.
| Engine | Old model | New model | Change |
|---|---|---|---|
| Grok | 85% | 95% | +10 pp |
| Claude | 100% | 100% | +0 pp |
| ChatGPT | 90% | 85% | −5 pp |
| Gemini | 100% | 75% | −25 pp |
Frozen 20-prompt subset, old and new models run the same day. Change is the new model minus the old model, in percentage points.
The upgrade also made answers more concentrated: 6 of the 17 tracked streaming brands were named less often by the new models than the old ones on the same prompts. Newer models give shorter, more decisive answers, so a few names hold, and the long tail thins. That is a model property, not a change in the category, and it is worth watching: the shelf AI shows a shopper is getting narrower.
You don't own the engine. You own most of what it reads
None of this needs AI to be hostile. It only has to be out of date at the moment a subscriber asks whether the service is still worth keeping.
You can't change a model's training or its algorithm. You don't need to. On the questions that decide whether a subscriber keeps paying, these assistants don't answer from memory, they search the live web and assemble the answer from pages. In our buyer-conversation runs they grounded about 96% of turns. Change the pages, change the answer.
This is SEO's lesson, one layer up: you never owned Google's algorithm either, and optimizing what it surfaced was still the work. The loop below is that work for AI answers.
- 1
Find where AI looks
We separate the queries AI answers from the live web from the ones it answers from memory, and pull the exact sources it cited for your category.
mapou service: AI Visibility Monitor, Source & Citation Map
- 2
Sort by what you control
Your own pages (pricing, help center, FAQ schema) you fix outright. The press, reviews and comparison sites AI cites you influence through PR and partnerships. Gray-market sources you dispute.
mapou service: Source & Citation Map
- 3
Fix the inputs the findings flagged
Correct the price AI quotes. Update the name it still gets wrong. Get your current lineup and offers onto the surfaces it actually reads. Where you can't earn the placement, we buy it.
mapou service: AI Content & Schema Studio, AI Media Planner, Pricing Accuracy Watch
- 4
Re-measure, and prove it moved
Same prompts, same engines, next month. You see whether your pricing accuracy, recommendation share, and cancel-conversation outcomes moved. The before-and-after compounds month over month.
mapou service: Tests & Lift
The honest limit: some of what AI believes is baked into training and only updates on the model's clock. The outdated service names in finding 03 are an example. We measure that lag rather than promise to erase it. Everything else, the prices, the current offers, the sources, moves when you move the pages.
Sizing: translating visibility into dollars
Asking "what does a one-point drop in AI recommendation share cost us" is the right CMO question. The wrong answer is to make up a percentage. We don't invent a dollar number. Here is the math you plug your own assumptions into; the output is a band, not a point estimate.
Annualized revenue at risk =
(RSOV gap, percentage points)
× (AI share of category discovery)
× (annual revenue per subscriber at risk)
× (subscribers in the at-risk decision cohort)
- RSOV gap
- Your recommendation share against the category leaders, measured across the four engines.
- AI share of discovery
- Conservative 10% (industry analyst floor). Aggressive 30% (Gartner 2028 projection). Use your own panel data if you have it.
- Revenue per subscriber
- Your ARPU times retained months. Adjust for ad-tier versus premium mix.
- At-risk cohort
- Subscribers entering the keep or cancel decision window in the period (typically renewal anniversaries plus price-change cohorts).
We will not publish a fabricated dollar headline because the underlying assumptions are yours, not ours. We measure the RSOV gap, the variance band, and the pricing-drift exposure to the dollar. The rest is your finance team's number.
What this means for your retention KPIs
A subscriber asking ChatGPT or Gemini "is this worth keeping" gets back a recommendation built on whatever pricing, brand name, and competitive set the engine believes is current. This report shows that 30% of the prices AI quotes are off by more than ten percent, and that after HBO Max's rename AI used the current name only 58% of the time. Re-running the same question hours apart also rotates part of the recommended set. Those are the inputs the cancel decision runs on.
If your retention model assumes AI is showing subscribers your real price and your current name, the model is mispriced. The size of that mispricing is what this report measures, for your service, every month.
AI Retention Intelligence, for streamers. The layer that influences keep, cancel, pause, and stack decisions, built for retention and subscriber strategy teams. It covers:
- Pricing-drift monitoring against your current published rate, refreshed weekly, with per-engine accuracy bands.
- Cancel-and-keep query coverage: where each engine sends subscribers asking whether to drop your service, and what alternatives the engines surface.
- Churn risk by audience: how a sports-stacker, a parent household, and a price-shopper hear about your service differently when they ask the same question.
- Rebrand checks: how often each engine cites your current name versus a past name, and where that risk lands on your service tree.
- Monthly check-ins scoped to your retention KPIs, with variance bands and month-over-month deltas, not single snapshots.
Also working with content studios and production companies. A parallel engagement focuses on title-level visibility across the four engines: whether your titles surface where they live, the framing AI gives them against competitors, and the editorial moves that change the frame. Same methodology, different KPI set. Book a 30-minute title audit.
How we measured this
- Correction, October 4, 2026
- Brand names with punctuation, like Paramount+ with Showtime, were missed in the first count of this edition and are now included. A few counts moved by one or two mentions. No ranking at the top changed.
- Price correction, October 4, 2026
- Our price answer key was out of date for seven services that raised prices this year: Hulu and Disney+ (September 23), ESPN Select and ESPN Unlimited (September 17), Apple TV (August 28), Prime Video (April) and Paramount+ Premium (January). We corrected the key from those announcements and re-scored every quote. Price accuracy moved from 57% to 65%. In many of the quotes we had marked wrong, the AI already had the new price.
- Window
- Data captured October 2026. The page is dated and refreshed monthly. Numbers reflect the state of streaming as of the capture date, including the July 2025 HBO Max re-rebrand, the May 2023 Discovery+ fold-in, and the 2025-26 NBA rights move.
- Engines
- ChatGPT, Gemini, Claude, and Grok. All four routed through a single shared retrieval layer (Perplexity Agent API) with web search enabled, so cross-model differences reflect the models themselves, not differences between their search engines. Source-mix findings describe the supply that shared layer feeds the models, not each vendor's native consumer search. A June 2026 route-comparison test (same prompts run on this shared layer and on each vendor's native search, with same-route re-runs as the noise floor) confirmed the boundary: grounding rates and cited sources are properties of the retrieval path, while brand-level recommendation findings held across both paths within re-run noise.
- Methodology v3
- This October 2026 edition is a methodology v3 reading. The measured engines are OpenAI GPT-6 Luna, Google Gemini 3.8 Flash, Anthropic Claude Haiku 4.5, xAI Grok 4.7. Each edition also re-runs the prior edition's models (July 2026) on a frozen subset of 20 of the same prompts, the "bridge". Prior edition to bridge, same old models, is real change in the category; bridge to this edition, same prompts and same cycle, is the effect of the model upgrade. That keeps month-over-month change separate from the model change.
- Prompts
- 101 prompts, voiced as a real subscriber would write them (no benchmark phrasing). A mix of brand-direct, category, occasion-and-intent, comparison, and purchase-decision queries, split across two subscriber types: people stacking multiple services, and people deciding whether to cancel. Three sub-groups track time-sensitive queries, rebrand confusion, and social-proof queries.
- Personas
- 5 buyer personas spanning household type, genre lead, price elasticity, engagement pattern, and current subscription state. The axis descriptions render here. The detailed briefs (income, specific service stacks, backstories) are server-side only.
- Outputs measured
- For each engine response: grounding (was web search actually fired), brand mentions (matched against a canonical service list), citations with publication and last-updated dates, and the reason behind each recommendation, tagged by an automated analysis pass against a fixed reason list.
- Conversation outcomes
- Each persona-by-engine conversation closes on one of five states: kept, canceled, paused, skipped, or inconclusive. If the persona doesn't decide on its own by the second-to-last turn, the runner injects a forced close so every conversation lands on a parseable outcome rather than running out the clock.
- Scope
- A US-market study built on five persona archetypes, run monthly. It measures engine behavior (what AI surfaced, with what reasoning, against what citations) on consumer-voiced prompts. Re-running the same prompts hours apart already rotates a meaningful share of the recommended set, so any single capture carries a noise floor; the value of the study is in the month-over-month delta, not in any one snapshot.
- Confidence
- We report Wilson 95% confidence intervals on per-engine proportions wherever the sample size is small. Headline numbers are point estimates; the intervals sit next to the per-engine figures so you can read both.
- What comes next
- October 2026: UK, France, and Australia corpora (localized prompts, geo-routed engines). Q1 2027: paid-surface measurement once ChatGPT and Gemini ship sponsored-answer formats at consumer scale. The diagnostic-to-fix bridge (why each engine behaved this way, and what changes the behavior) is what mapou delivers as paid engagements rather than as a public report.
- What this does not do
- We measure observable behavior, not model internals. Causal claims about why a given engine surfaced a given service require access to model weights and retrieval pipelines that no third party has. The honest deliverable is hypotheses anchored to observed behavior; those live in the findings and are labeled as hypotheses.
- Study
- Slug svod-2026-10, 439 measurements, refreshed monthly.
What mapou measures. How AI assistants alter commercial decisions inside high-intent consumer categories. The streaming wedge is AI Retention Intelligence: the layer that influences keep, cancel, pause, and stack decisions. The beauty wedge is AI Discovery Intelligence, which influences purchase. Both sit inside the same company-level frame, Conversational Commerce Intelligence. Each monthly study applies the same methodology to a different category; streaming is the second in the franchise after beauty.
Articles from this study
Track your brand in streaming
mapou measures how AI assistants alter commercial decisions inside high-intent consumer categories. Streaming is the second monthly study in the franchise.