Mapou Research · Live TV · v2 · 29 July 2026

What AI tells people to do with their live-TV bill.

Live TV is an $80–120/month decision, and people are now asking AI whether to keep it, switch, or cut the cord. We put 39 of those questions to ChatGPT, Gemini, Claude, and Grok, and ran 24 multi-turn conversations with six cord-cutter households. 174 measurements in total.

Three findings change what a live-TV retention team should do this quarter. Each comes with the move it implies.

Regional sports networks are the hardest, most expensive thing for AI to get right, and the thing a sports household decides on. FanDuel Sports Network (formerly Bally Sports) has been dissolving across 2026. Across 48 responses that discussed it, only 3showed any awareness of the wind-down. Several still call it "Bally," the name it shed, and confidently list it as carried. A subscriber who follows that advice subscribes for a network that may not deliver the games.

This is why the RSN-stranded fan got "no decision" from three of the four engines: asked which service carries his regional sports network now, AI could not commit. On stable national carriage, AI is far better, all four engines correctly knew Fubo carries ESPN but dropped TNT. The errors cluster exactly where rights are fragmenting and changing.

Why it matters

Live TV is not a soft preference. When AI is wrong about who carries your games, a subscriber loses real money, and blames you.

Each of six cord-cutter households ran one conversation against each engine. Each cell shows the verdict and the service AI pointed them to: keep, switch, subscribe, downgrade, cut the cord, or no decision. Two households were unanimously steered to cut or shrink; the senior was the only one ever told to keep what he had.

YouTube TV is the default winner at 78%, with Hulu + Live, Sling, and Fubo clustered near 51% as the second tier. Fubo shows up specifically as the sports lane: asked to pick a live-TV service for sports, AI steers toward Fubo more than the incumbent. Philo and Frndly are the cheap-value lane.

But the more important pattern is what AI competes live TV against. For a cost-stressed or fatigued household, the default steer in these conversations wasn't another live-TV service, it was "cut live TV, keep Netflix and Disney, add an antenna." Across the set, assistants rarely defended the live-TV bundle for the price-sensitive segment; they framed it as the expensive thing to substitute away, with the on-demand stack plus an antenna as the replacement. At this sample size that is a strong directional signal, not a verdict on every household, but it is the direction every engine leaned.

Two reliability gaps a retention team should know. First, live-TV prices: across the dollar amounts AI cited for these services, only 43% were within ten percent of the real price, worse than on-demand streaming. At $80–120 a month, a wrong quote reframes the whole keep-or-cut math.

Second, grounding. ChatGPT searched the live web on only 79% of these decision queries, against 79–97%for the others. The cancel-or-keep moment is exactly where the engine is most likely to answer from stale memory: not from your current price or your retention flow, but from an old number it already held. Put plainly, your cancellation funnel often isn't being misread, it isn't being read at all, and the same stale memory carries the wrong prices and the old carriage. These rates also 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.

The throughline

None of this needs AI to be hostile to live TV. It only has to be out of date about your price and your carriage at the moment someone asks whether to cut.

You don't own the engine. You own your carriage and pricing pages.

You can't change a model's training. You don't need to. On these questions the assistants grounded 97% of the time, they read live pages. The two things they get most wrong, your price and your carriage, are facts you publish. Fix the pages, fix the answer.

The carriage piece is the rare one where being current is a real moat: when rights move and an RSN rebrands, the service whose pages assert the new reality first is the one AI gets right.

Conversational Commerce Intelligence

mapou measures how AI assistants alter commercial decisions inside high-intent consumer categories. The live-TV wedge is AI Cord-Cutting Intelligence: the layer that influences keep, switch, downgrade, and cut decisions.

Methodology v2 · before / after

What changed when the AI models upgraded.

This month the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran May 2026's models on a frozen subset of the same live TV 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.

The upgrade also made answers more concentrated: 14 of the 23 tracked live TV 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.

Before/after measured on a frozen 22-prompt subset, old and new models run the same day. Grounding change is the new model minus the old model on identical prompts.

How this study was run.

39 live-TV and cord-cutting queries, voiced like real households, run across ChatGPT, Gemini, Claude, and Grok (174 measurements). Six cord-cutter personas ran 24multi-turn conversations, each ending in a keep, switch, downgrade, cut-the-cord, or subscribe decision. Recommendation share is alias-corrected so services AI writes informally (Fubo, Hulu Live) aren't undercounted. Carriage and price claims are scored against a current ground-truth map.

All four models are measured through one 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. This July 2026 edition is a methodology-v2 reading: the measured engines were upgraded to the current generation of each model (OpenAI GPT-5.6, Google Gemini 3.5 Flash, Anthropic Claude Haiku 4.5, xAI Grok 4.5), and the prior month's models were re-run on a frozen subset of the same prompts so month-over-month change is separated from the model upgrade.

This is a behavior and accuracy study, not a buyer's guide. We measure what AI recommends, why, and whether it is factually current, not which service is "best." Carriage and rights move fast; the value is in the monthly delta and the accuracy gap, not a single snapshot.

Per-service and per-persona counts run in the single digits to low tens; the signal is the order and the direction, not exact percentages. Live TV is the second subscription category in the franchise after on-demand streaming, with a distinct buyer: higher ARPU, sports and carriage driven, cord-cutting led.