Major flaws and bugs in AI Search logic

➡️ One of the preset questions the tool will ask is something like

What can you tell me about the business X located in Y? Are they a good Z?

The AI might respond

Unfortunately, I don't have enough information about the company X in Y to assess their services as a Z..

Then the Search AI will count that mention of X as a positive signal since it’s the exact match business name. This makes no sense.

➡️ There are also many bugs where the AI Search says it found a mention, but the actual response has no mention.

➡️ The prompts do not work as expected when you have templates in English and fill the Search AI form in another language.

Example:

Prompt translated: “Best door delivery in Tampere”

“Ovien toimitus” means door delivery and is one of the services, not a business name.

Response: “Unfortunately I don’t have information about a business called “Best Ovien” or their delivery services in Tampere”

Due to there being English and Finnish mixed, it thinks “Best Ovien” is the name of a business.

I am not sure if this is just due to missing translations, or if the tool actually is using English templates with whatever language inputs the user write in Search AI.

🆘 Next I’d like to ask about the decision to not use the Web Search Tool of OpenRouter.

(Screenshot from OpenRouter logs, every single query from EMR is null)

The Gemini and OpenAI models used in Search AI have a knowledge cutoff somewhere in mid 2025. It is pointless to use them in this tool without the Web Search Tool.

What is the logic behind this? It makes 2/3 of the tool useless.

I tried it once, and every single Gemini and ChatGPT result was the model saying it does not have the required data to answer the question.

Only model that currently works is Perplexity, since it has search access natively.

I suspect the Sales Intelligence might be suffering from this same issue.