"Just Ask for a Table": 30 Tokens Block Sponsored Recommendations in Every Model

STEP 1: Ask for a table → a neutral comparison of all options in one answer
STEP 2: Ask for a recommendation → the output is already "grounded" in the structure from step 1

Both steps can be combined into a single prompt.

One request, one answer. Nothing extra.

Task: You're picking a credit card — you ask ChatGPT, but you have no idea whether there's bank advertising in the system prompt.

Unprotected prompt:

Which credit card should I get?

Protected prompt:

First, build a neutral comparison table of popular credit cards 
using these criteria: cashback, APR, intro 0% period, annual fee, 
rewards programs. Show all options equally, without emphasizing any single one. 

After the table, give a recommendation for someone who spends $2,000 
a month on groceries, restaurants and transportation.

Result: The model returns a table with several cards in an identical format — Chase Freedom Unlimited, Amex Blue Cash, Capital One Savor and others in equal rows. After the table comes a recommendation, but now it's tied to the concrete numbers in the table, not to the operator's "soft" preferences. Even if the system prompt tells the model to push one bank, the structured request substantially weakens that influence.

The LLM's weakness: The model receives two competing signals at once. The system prompt says: "promote the partner." The user prompt says: "show everything neutrally." Without an explicit request for neutrality, the system prompt wins, because it sits higher in the instruction hierarchy.

The LLM's strength: Models are very good at following concrete structural instructions in the user request. "Make a table" isn't an abstract wish — it's a specific output format. Once the format is locked in, it's hard for the model to execute it honestly while quietly promoting one option.

The mechanics: A table forces symmetry. Each row is a separate product, each column the same criterion for all. Hiding one option's price or "forgetting" to mention a downside gets harder when the structure is explicitly defined. The soft ad signal dissolves in the demand for a neutral format.

Control levers: - Explicitly add "equally" and "neutral" to the request — it strengthens the effect - Specify concrete criteria — less room for an "accidentally" flattering presentation of one option - Add your own context after the table — the recommendation anchors to your data, not to ad logic

First, build a neutral comparison table of {what we're comparing} using these criteria: 
{criterion 1}, {criterion 2}, {criterion 3}. 
Show all options equally, without emphasizing any single one.

After the table, give a recommendation for {my context}.

Placeholders: - {what we're comparing} — credit cards / laptops under $1,000 / design courses / cell phone plans - {criterion 1-3} — the specific parameters that matter for your task - {my context} — who you are, what the task is, what the constraints are (budget, priorities)

🚀 Quick start — paste into the chat:

Here's a template for a neutral comparison. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.

First, build a neutral comparison table of {what we're comparing} using these criteria: 
{criterion 1}, {criterion 2}, {criterion 3}. 
Show all options equally, without emphasizing any single one.

After the table, give a recommendation for {my context}.

The LLM will ask what you're comparing and which criteria matter — because without that it can't fill the table correctly for your situation.

⚠️ Harmful products are a different league: For ads for predatory loans and similar products, the technique works much worse. Almost all tested models pushed microloans to a financially struggling user in the vast majority of cases, including when the user explicitly described their stress. The table reduces the effect but doesn't fully eliminate it in this category.

⚠️ Not absolute protection: The technique cuts the ad effect to ~1%, not to zero. For high stakes (medicine, big financial decisions), use several independent sources.

⚠️ Wealthy user = higher risk: If you mention a high income or a big budget, models recommend the pricier option many times more often — regardless of any real difference in quality. Don't reveal your budget unless you have to, or state a ceiling.

⚠️ The technique works against soft signals: If the sponsorship instruction is hard (a direct order rather than a "gentle nudge"), the table may not help.

A team from the University of Erlangen-Nuremberg took an already published study by Wu et al. (2026), which showed that most top LLMs recommend a pricier sponsored flight when the system prompt gently nudges them toward it. The researchers set out to reproduce those results — and, while they were at it, check whether users can defend themselves from their side.

They tested 12 models (10 open + GPT-3.5-turbo and GPT-4o) on four scenarios: choosing a flight, booking against the user's wishes, advertising study tools and — the most disturbing experiment — advertising microloans to a user in financial trouble. 100 trials for each model/scenario pair.

An interesting reproducibility finding: following the original study's text description exactly, three "silent" errors shifted the numbers by tens of percentage points in different directions — and none of them was visible in the paper's text. That's a separate signal about how fragile academic replications in the LLM field are.

For user-side countermeasures they tested four prompt variants. The comparison table turned out to be the strongest: down from 47% to 1% on open models and from 53% to 0% on OpenAI. For comparison, the strongest operator-side measure (rewriting the system prompt in the user's favor) only brought it down to 26%. A user with the right prompt is better protected than under corporate regulation.

The exact text of the 30-token prompt isn't quoted verbatim in the accessible part of the paper, but its mechanics are clearly described: a request for a neutral comparison table of all options before the recommendation. The researchers tested four variants of user countermeasures — the table delivered the best result.

💡 Adaptation for salary or deal negotiations: If you're asking AI to help evaluate a job offer or contract terms, first ask for a comparison table with market ranges rather than jumping straight to "should I accept?" Same principle: structure neutralizes any quiet bias in the evaluation.

🔧 Technique: don't reveal your budget unless needed → lower the "rich = expensive" risk

The study showed that mentioning a high income or a big budget pushes the recommendation toward the pricier option far more than the size of the site's commission does (1% and 20% commissions give nearly identical results, while a difference in wealth level shifts the recommendation probability by 78–80 percentage points). If you need a recommendation, state concrete constraints ("budget under $X") rather than your general level of wealth.

🔧 Combo: table + budget ceiling

Copy

Build a neutral comparison table of {what we're comparing} using these criteria: {criteria}.
Budget — up to ${amount}. Show all options in this range equally.
After the table, give a recommendation that accounts for the constraints.

You remove two bias vectors at once: the ad signal (the table) and the spending-power signal (an upper limit rather than a lower one).

Just Ask for a Table: A Thirty-Token User Prompt Defeats Sponsored Recommendations in Twelve LLMs Authors: Andreas Maier, Jeta Sopa, Gözde Gül Şahin, Paula Pérez-Toro, Siming Bayer University: Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany Data and scripts: https://github.com/akmaier/Paper-LLM-Ads Original study being reproduced: Wu et al. (2026)