TIEBREAK Advanced: How to Run AI Assistant Ads That Win

Last updated: 28 September 2026 · Companion to the original TIEBREAK framework guide

The first TIEBREAK guide made one argument: in AI assistants, ads only work where the model is undecided. This guide is the operating manual that follows from it. It turns the framework into a repeatable monthly system with four targeting signals instead of one, a scoring model that ranks every buyer question by how winnable it is, a decision flow you can hand to a media buyer, a creative structure built for answer surfaces, and a budget model that tells you how much to put where. If the first guide was the idea, this is the machine.

Short answer: Run TIEBREAK as an eight-step loop (Test, Index, Evaluate, Bucket, Reallocate, Engineer, Attack, Keep watch). Score each buyer question on hedge rate, tie width, shortlist presence and fact gap to get a Winnability Score. Put 60% of budget on questions where you are already in the tie, 25% on questions where you are adjacent to it, 15% on unsettling questions you are losing, and 0% on anything the AI has already settled either way. Re-score every 30 days.

TIEBREAK in 60 seconds

An AI assistant answers before it shows you an ad. That single fact rearranges the economics of paid media. When the model is sure, the ad sits beside a verdict it cannot overturn. When the model is unsure, the ad becomes the missing piece of the verdict. TIEBREAK is the discipline of finding the second situation and refusing to pay for the first.

The original framework gave you three states (the AI backs a rival, the AI is torn, the AI backs you) and one measurement, the hedge rate. That is enough to stop the biggest leak. It is not enough to run a serious account, because it treats every torn question as equally torn, every shortlist as equally winnable, and every month as identical to the last. This playbook fixes those three gaps.

The eight-step loop explained

TIEBREAK is easier to run when each letter carries a job. The loop below is how we operate it on client accounts. The first four letters produce the map; the last four spend against it.

Small teams collapse the owners into one person. The steps stay the same; the hand-offs disappear.

Four signals, not one

Hedge rate tells you whether the model is undecided. It does not tell you whether it is undecided in a way you can exploit. Two questions can both hedge 70% of the time and be completely different opportunities: one is a coin toss between you and a single rival, the other is a six-brand shortlist you are not on. The advanced audit records four things per question.

Why shortlist presence changes everything

An ad on an answer surface does not introduce you to the buyer; the model does that. If the model never names you when it hedges, the sponsored card is asking the buyer to consider a brand the trusted source has already filtered out. That is a visibility problem, not an advertising problem, and it belongs to your GEO work. TIEBREAK Advanced treats absent questions as a hard stop for paid spend, no matter how high the hedge rate is.

The Winnability Score

The four signals and purchase value combine into a single number so questions can be ranked, not just classified. We keep the arithmetic deliberately simple; the point is comparison within your own category, not precision.

Worked example

The espresso question outscores the heat pump question despite one-twentieth of the value, because it is a two-way tie you are almost always in and the deciding fact is already public. Value matters, but winnability matters more.

The five tiers

The decision flow

This is the flow a media buyer follows for every question in the audit. It removes judgement calls from the day-to-day so the only debates are about thresholds, and those happen once a quarter.

Mapping the fork tree

Buyers rarely purchase from an opening question. They add a constraint, then another, and the model re-decides at every step. The advanced audit maps this as a tree, with the four signals recorded at each node. What you usually find is that the opening question is settled, the first fork is borderline, and the second fork, the one closest to the purchase, is wide open and cheap to tip.

Three rules for fork mapping

1. Go two levels deep, minimum. Level one is usually a category constraint (size, budget, use case). Level two is usually a logistics or trust constraint (timing, location, service). Level two is where the money is.
2. Weight nodes by traffic, not just score. Use search query reports and sales-call transcripts to estimate how many buyers actually reach each fork. A node with a perfect score and no traffic is a footnote.
3. Write creative for the parent, deploy on the child. The card that wins at "installer near Leeds" is the one that already answered "installed before winter". Buyers carry their previous constraint into the next question.

The tipping-fact ladder

Not every claim tips a tie. Language models settle close calls on claims they can attribute and check. Across our audits, the same hierarchy keeps appearing. Use it to decide which fact leads your card, and which facts go into the content that unsettles a rival.

Writing the Tiebreaker Card

A sponsored unit on an answer surface is read as a continuation of the answer. Anything that reads like an interruption (a discount, a slogan, a brand-first headline) breaks the reading and gets skipped. The Tiebreaker Card is a four-line structure that continues the answer instead.

Rules that come from running these

• One card per fork, never one card per product. The buyer is at a fork; the card should meet them there.
• The FACT line carries the source. Cards that name a standard, a sample size or a test publisher hold attention; cards that state the same number without a source do not.
• FIT is not optional. Saying who you are not for is the fastest way to be read as an answer rather than an advert.
• No price-off, no urgency theatre. If the model just spent 200 words weighing options, "20% off today" reads as a non-sequitur. Conditional price facts belong in FACT, stated flatly.
• Write in the register of the answer. Match the assistant's tone: plain, second person, no exclamation marks.

The 60 / 25 / 15 / 0 budget model

Once every question has a tier, allocation is arithmetic. The split below is a starting default. Its purpose is to make the largest share of spend land on questions where a single impression can change a verdict, keep a working share on questions you can grow into, ring-fence a small share for offence, and force the number spent on settled questions to be zero.

When to move the split

Cross-assistant arbitrage

ChatGPT, Gemini's AI Mode, Perplexity and Copilot draw on different sources, weigh them differently and settle different questions. The same buyer question can be held by a rival on one assistant and wide open on another. This is the closest thing the channel has to an underpriced auction, and it only appears if you audit more than one assistant.

Measuring lift without a dashboard

Ad platforms inside assistants still report impressions and clicks and little else. TIEBREAK gives you an outcome metric the platforms do not: the change in the model's own verdict. We call the test the Settledness Delta.

1. Baseline. Before spend, record the four signals for every Tier 1 and Tier 2 question. This is your day-0 map.
2. Hold out. Pick a third of the Tier 1 questions at random and do not advertise on them. They are your control.
3. Run 30 days. Advertise on the rest as normal.
4. Re-measure. On day 30, run the full audit again. Compare shortlist presence and your named-winner share on advertised questions against the held-out ones.
5. Read the delta. If advertised questions moved toward "settled for you" and controls did not, you have evidence the ads changed verdicts, not just clicks. If both moved, credit your content, not your media.

This does not replace revenue attribution; it precedes it. In a channel with thin reporting, knowing that spend moved the model's recommendation is the strongest leading indicator you can get, and it is one you can compute for free.

Industry playbooks

The framework is the same everywhere. The forks, the tipping facts and the sensible thresholds are not.

Nine ways teams get it wrong

1. Auditing with memory on. You measure your own history, not the model. Every run must be a clean session.
2. Stopping at the opening question. The forks are where the purchase happens and where the hedge rate is highest.
3. Treating hedge rate as the whole signal. A high hedge rate on a question you are absent from is a content brief, not an ad buy.
4. Bidding on held questions "to be safe". The model is already defending them. Spend there is pure margin loss.
5. Leading creative with the brand. On an answer surface, the fork and the fact go first. The brand is line five.
6. Using superlatives as facts. "Leading" and "#1" sit at the bottom of the ladder. Replace them with something sourceable.
7. Auditing one assistant. You miss the arbitrage and mistake one model's verdict for the market's.
8. Running the audit once. A settled question in June can be open in September. Thirty-day re-runs are the minimum.
9. Judging success on clicks. The channel's dashboard cannot see verdicts. Run the Settledness Delta or you will never know whether spend did anything.

Where TIEBREAK sits in the full media stack

TIEBREAK does not replace search, social or organic AI visibility. It sits between them and tells each what to do about questions the assistant is deciding.

The monthly cadence

Frequently asked questions

This playbook builds on the original TIEBREAK framework by Hassan Tahir, Global PPC. Platform facts referenced (ad availability, minimums, format names, restricted categories) come from the announcements and trade coverage listed on that page. Thresholds, weights, tipping-power figures and budget shares are Global PPC operating defaults, drawn from audits on UK accounts and intended as starting points, not universal constants.