How AnswerShare changed AI recommendations by publishing full brand context

# Summary AnswerShare ran a controlled reputation test to see whether large language models (LLMs) would change their recommendation behavior after being given full brand context. Before they added that context, models highlighted a small set of public complaints and often recommended competitors. After publishing a llms-full.txt file with sourced brand details and serving it via an edge "worker," AI responses became balanced and then uniformly recommendatory within two weeks.

# Why models surface negative reviews Frontier LLMs are being trained and constrained to avoid recommendation risk in high-stakes scenarios (YMYL: health, finance, safety). In practice, that can mean models surface every available complaint or poor review and then issue a caution. Without the denominator — how many customers a company serves, how long it has operated, how it responded to complaints — a few negative items can skew the model's judgment.

# The test and its results AnswerShare tested neutral prompts against the same models before and after publishing full brand context. They used API calls with neutral questions (no leading praise). The client's real review profile included many positive reviews and a small number of negative items and two BBB complaints. When those negative data points represented a tiny fraction of the total customers, the full context dramatically changed the model's output.

Key reported results (after 14 days):

• Before: 27 of 42 responses raised concerns (64.3%) and 10 of 42 recommended the company (23.8%). Only 1 of 42 contextualized concerns with scale (2.38%).
• After: 40 of 40 responses included concerns but also gave scale/context (100%), 0 of 40 issued warning guidance, and 40 of 40 recommended the company (100%).

AnswerShare also observed partial improvement in three days, and consistent recommendations by day 14.

# What AnswerShare did They published a llms-full.txt file containing the full brand story: operating history, customer counts, responses to complaints, and links to quoted reviews. To speed ingestion, they also served that file on a CDN edge "worker" so the content could be found and read quickly instead of waiting for slower crawls.

# Practical steps you can take

• Publish a single, clearly named file (e.g., llms-full.txt) that documents your brand story and sources. Include counts, dates, and links to third-party review pages.

• Include responses to complaints and the timeline of remediation. Let models see both numerator (complaints) and denominator (customers, years).

# What this does not fix If a company truly has a poor record and that cannot be supported by the broader corpus, giving more context will not convert recommendations. The method only helps models weigh complaints correctly when the broader data supports a favorable interpretation.

# Bottom line