TIEBREAK Framework: Only Pay for AI Ads When the AI Can’t Decide

Last updated: 25 September 2026 · Updated as ChatGPT, Gemini and Perplexity ad formats change

Ads now run inside ChatGPT, Google’s AI Mode, Perplexity and Copilot. Most advertisers are buying them the way they bought Google keywords a decade ago: by intent. That is the wrong variable. The TIEBREAK framework is a new model for AI assistant advertising built on one observation: an ad cannot beat an answer the AI is already sure about. It only changes the outcome when the AI is torn. This guide explains the framework in full, the research behind it, how to run it on your own account, when it works, and when it doesn’t.

Short answer: TIEBREAK is a decision framework for paid ads in AI assistants. It sorts every buyer question into three states (the AI already picks a rival, the AI can’t decide, or the AI already picks you) and tells you to spend only where the AI can’t decide. The targeting signal is the hedge rate: how often the assistant gives no single winner when asked the same question repeatedly. High hedge rate plus high purchase value is where ChatGPT ads budget belongs.

What is the TIEBREAK framework?

TIEBREAK is a marketing framework for paid advertising inside AI assistants. It was developed at Global PPC in September 2026 as we started managing ChatGPT ads for UK clients and noticed that the playbook we had used for Google Ads and Meta Ads for seven years did not transfer.

The core claim is simple. In social feeds, the scarce resource was attention. In search, it was intent. In AI assistants, it is indecision. A search results page shows ten links and lets the user choose. An AI assistant gives one recommendation. If that recommendation is confident, an ad rendered beneath it is sitting next to an answer that has just told the user to buy something else. The ad can be seen, but it cannot win.

TIEBREAK therefore treats the assistant’s confidence as the targeting variable, not the user’s intent. It gives you a way to measure that confidence with nothing more than the assistant itself, a spreadsheet and an afternoon, and it tells you to move budget away from questions the AI has already settled and towards questions where the AI is torn between options.

Like Seth Godin’s Purple Cow or the Bullseye framework from Traction, TIEBREAK is a lens rather than a tool. There is no software to buy. It is a way of deciding where money should and should not go in a channel that does not yet report the number that decides the sale.

Why intent targeting fails in AI assistants

Every advertiser who has run Google Ads has been trained to buy intent. A search for “espresso machine under £600” signals a buyer, so you bid on it. That logic is sound when the search engine shows a page of options, because the user is still deciding and your ad is one of the options on the table.

An AI assistant breaks that logic in three ways.

• The assistant decides before the ad renders. In ChatGPT the sponsored unit appears beneath the answer, and OpenAI has stated plainly that ads do not influence the answers the model gives. The recommendation is written first; the ad is attached afterwards. If the answer says “buy the Sage Bambino”, an ad for a different machine is arguing with a source the user just trusted.
• You are buying a context, not a person. OpenAI’s ads run on contextual matching, and advertisers get aggregate views and clicks rather than any view of the individual. You cannot out-target a competitor with audience data the way you can on Meta. The only lever left is what the model says.
• The scale is real, so the waste is real. ChatGPT is used by roughly three-quarters of the estimated 153 million US generative AI users, and OpenAI moved its ads pilot from a $200,000 minimum to a $50,000 self-serve minimum in April 2026. Money is flowing into a channel where the most important variable is invisible in the dashboard.

Put those together and the conclusion is uncomfortable: a large share of what advertisers are currently buying in ChatGPT, the high-intent commercial questions, is partly wasted, because the model has already made up its mind before the sponsored card appears.

The research behind TIEBREAK

TIEBREAK rests on three sets of evidence: what the platforms have shipped, how large language models behave, and what early advertisers have reported. Here is each in turn, so you can check the reasoning rather than take it on trust.

1. What the AI ad platforms have actually shipped

2. How large language models actually answer

The framework depends on a property of language models that anyone can verify: the same question does not always produce the same answer. Models sample from a probability distribution, so a question with one dominant answer returns that answer almost every time, while a question with several close candidates returns different picks on different runs, or hedges with “it depends” and a shortlist.

That variance is not noise. It is a direct readout of how settled the model’s recommendation is. A question that produces the same brand 19 times out of 20 is settled. A question that produces four different brands and six shortlists in 20 runs is open. TIEBREAK simply turns that readout into a number and a budget rule.

Two further properties matter. First, models resolve close calls with facts they can source: a specific, verifiable claim (a test result, a spec, a price condition) moves a recommendation where a generic superlative does not. Second, persistent memory now shapes answers per user: OpenAI’s memory feature has been live since 2024 and expanded in 2025, and the same question can return a different shortlist for each person. This is why TIEBREAK measures across fresh sessions and multiple assistants rather than trusting a single run.

3. What early ChatGPT advertisers have reported

Trade coverage through spring 2026 is consistent on one point: measurement is thin. Early advertisers received little beyond impressions, which made it hard to tie spend to business outcomes, and OpenAI has been addressing this with click-based pricing and a LiveRamp partnership. Creative guidance is equally consistent: units that open with a discount code collapse in engagement, because the surrounding text is an answer and a discount reads as noise.

Both findings point the same way. If the dashboard cannot tell you whether an ad changed the outcome, and the creative only works when it completes the answer rather than competing with it, then the question you need to answer before spending is: was there an outcome to change? That is the question TIEBREAK is built to answer.

The three question states

TIEBREAK classifies every buyer question in your category into one of three states, based on how the assistant answers it over repeated runs. Each state has one instruction.

Classification thresholds we use: settled = one brand named on 70%+ of runs; open = no brand above 50% or an explicit shortlist on 40%+ of runs.

Settled against you: why you don’t bid

This is the state advertisers find hardest to accept, because it is exactly the high-intent question they want to own. But an ad placed beneath a confident recommendation for a rival is not competing on equal terms. The user has just been told what to buy by a source they trust; your sponsored card is a stranger disagreeing. Engagement is low and, worse, the impression teaches the user to associate your brand with “the one the AI didn’t pick”.

The play here is to change the facts before you change the budget. Publish the head-to-head test, the spec the competitor cannot match, the price condition under which you win. Get it onto pages the model reads and cites. Then re-run the question. When the hedge rate climbs, the question has moved to “open”, and now your ad has something to tip.

Open: why this is the whole opportunity

An open question is one where the model genuinely cannot separate the candidates. Here the sponsored unit is not fighting the answer; it is completing it. A card that supplies the one fact the model was missing, such as “the only machine in this price band with a PID controller”, resolves the user’s indecision and the model’s at the same time. This is where a single ad impression changes a purchase, and it is why TIEBREAK concentrates budget so aggressively.

Settled for you: why you spend nothing

Paying to be recommended where you are already recommended is the most expensive mistake in this channel, and the easiest to make, because it feels like defending your territory. It isn’t. The model is already doing that for free. Your job is surveillance: re-run the question monthly, watch for a rising hedge rate, and treat that as an early warning that a competitor has published something the model finds persuasive.

Hedge rate: the metric that matters

The hedge rate is the share of runs in which the assistant gives no single winner. Ask “best project management tool for a 12-person team” 20 times in fresh sessions. If 14 of those answers name one product outright and 6 give a shortlist, the hedge rate is 30%. If 3 name a product and 17 hedge, it is 85%.

Plot every question in your category on two axes, hedge rate across and purchase value up, and you have the settledness map. Budget goes to the top-right: questions the AI cannot decide, about purchases worth real money.

The fork rule: decisions happen on question two

In search, the decision happens on the query. In chat, it happens on the follow-up. Almost nobody buys from the first answer to “best espresso machine”. They narrow: “under £600”, “for a small kitchen”, “for a beginner”. Every narrowing resets the tie. A question that was settled at the top level becomes open the moment the user adds a constraint the model has thin data on.

The fork rule says two things. First, map hedge rates on the follow-ups, not just the opening question; the follow-ups are usually more open and closer to purchase. Second, write your sponsored creative as a pre-answer to the three most common forks. The ad that wins is the one that has already resolved “but is it good for a beginner?” before the user asks it.

How to run TIEBREAK step by step

1. List the questions. Write down every question a buyer in your category might ask an assistant between “I have a problem” and “I bought it”. Include the forks. Aim for 20 to 40 questions. Your search query reports and sales call notes are the best source.
2. Run each one 20 times. Fresh session, memory off, no account history. Do it on ChatGPT and at least one of Gemini AI Mode or Perplexity. Record the named winner for each run, or “hedge” if there is no single winner.
3. Calculate hedge rate and share. Hedge rate is the share of runs with no single winner. Share is how often each brand is named. Together they tell you the state: settled for you, settled against you, or open.
4. Add purchase value. Attach an average order value or deal size to each question. Plot hedge rate against value. Circle the top-right.
5. Reallocate. Move ChatGPT ads budget from settled questions to open ones. Write the creative for the fork, not the opening question, and lead with the one provable fact that resolves the tie.
6. Unsettle where you’re losing. For high-value questions settled against you, publish the counter-fact, get it cited, and re-measure monthly. Bid only when the hedge rate rises.
7. Re-run every 30 days. Models update, competitors publish, and formats change. A question’s state is not permanent. Treat the audit as a monthly report, not a one-off.

When to use TIEBREAK, and when not to

Use it when

• You are spending, or about to spend, on ads inside an AI assistant: ChatGPT, Google AI Mode, Perplexity, Copilot or an independent prompt-triggered network. The framework exists for this inventory.
• Your category has real competition. TIEBREAK is a tool for deciding where to fight. If there are four credible options for what you sell, there are questions the AI cannot decide, and that is your opening.
• One sale is worth thousands. The audit takes real hours. It pays back fastest for high-ticket categories such as home energy, private dental, B2B SaaS and premium ecommerce, where flipping a single recommendation is worth the work.
• You already do GEO or AI visibility work. TIEBREAK is the paid-media half of that discipline. GEO changes what the model says; TIEBREAK decides where to pay once you know what it says.

Don’t use it when

• You have no competition in the assistant’s answer. If every question is settled for you, the audit will tell you to spend nothing. That is the correct answer, but you didn’t need a framework to reach it.
• Your category is blocked. OpenAI does not show ads near health, mental-health or political queries. Regulated clinics should treat AI Mode and ChatGPT as visibility channels, not paid ones, until policy changes.
• You are a one-person operation with a small budget. The self-serve minimum is $50,000. Below that, run the audit for intelligence and put the budget into Google Ads, where the auction still lets you win on a confident question.
• You want it to replace Google or Meta. It doesn’t. TIEBREAK is a rule for one channel. Search still rewards intent; Meta still rewards creative. Run them on their own logic.

Applying TIEBREAK to ChatGPT, AI Mode and Perplexity

The framework is platform-neutral, but the platforms are not identical. Here is how the audit and the budget rule translate to each.

For UK costs, targeting rules and creative requirements on the flagship platform, our ChatGPT Ads UK 2026 guide covers the mechanics. For the Google side, see our Google Ads updates tracker, which follows AI Mode ad formats as they expand.

Four new metrics for AI assistant advertising

None of these appear in an ad dashboard yet. That is the point: TIEBREAK is a framework for the gap between what is measurable today and what actually decides the sale.

The fourth is the settledness map itself: hedge rate against purchase value, refreshed monthly. Over time it becomes the single most useful picture of where your category is winnable, and it is a picture no competitor can copy without doing the work.

Frequently asked questions

Hassan Tahir

Founder & Lead PPC Strategist · Global Performance Propel Company · Google Partner & Meta Business Partner