Prism matching experiment
Can synonyms change whether an ad qualifies in AI chat?
In this six-question test, one question scored 0.436 without synonym matching and 0.806 with it. The ad stayed the same. Prism's eligibility threshold was 0.65.
What changed for this question?
“Turn a photo into a clip with AI” crossed the matching threshold when Prism treated “photo” as “image” and “clip” as “video.”
That is a 0.370 increase in the matching score. It is not a percentage increase in clicks, revenue or conversion. Clearing the threshold makes an ad eligible; it does not prove the ad was shown.
What were all six results?
We selected five video-related questions and one unrelated weather question for this illustration. This is a small, deliberately chosen test set, not a random sample of visitor questions. All results are below, including the scores that fell slightly.
| Test question | Without synonyms | With synonyms | With synonyms, at 0.65 |
|---|---|---|---|
| How do I make an AI video from a prompt? | 0.742 | 0.726 | Above threshold |
| Turn an image into a video with AI | 0.786 | 0.806 | Above threshold |
| Turn a photo into a clip with AI | 0.436 | 0.806 | Above threshold |
| Best AI video generator | 0.683 | 0.676 | Above threshold |
| AI video generator for short clips | 0.698 | 0.784 | Above threshold |
| What is the weather today? | 0.000 | 0.000 | Below threshold |
Four of six questions cleared 0.65 without synonyms; five of six cleared it with synonyms. The weather question scored zero in both runs. One unrelated question is not enough to establish a general false-match rate.
Which ad was tested?
The fixed example was the Higgsfield affiliate ad already in Prism's library. Higgsfield did not sponsor this test. The wording below was unchanged between runs.
Title: Make AI videos from a prompt or an image Description: Turn a text prompt or a still image into a short video with Higgsfield. Keywords: how do I make an ai video from a prompt, turn an image into a video with ai, best ai video generator, ai video generator for short clips
How was the comparison run?
On October 16, 2026, we ran Prism's matching code from version 015e321 twice against the same fixed ad and six questions. In the control, the synonym dictionary was empty. In the other run, the existing dictionary was enabled. The word stemming, score calculation and threshold were unchanged. Scores are rounded to three decimal places; threshold decisions use the unrounded scores.
This comparison tests the synonym layer within the current matcher. It is not a comparison of two complete historical deployments. It uses one question at a time, with no conversation history, competing ads or serving gates. No AI answer model was called, no live ad impressions were generated and no private visitor questions were used.
Download the complete test inputs and unrounded results (JSON).
What can an advertiser learn from this?
Buyers may describe the same task with different words. Testing those word choices can reveal where an ad might not surface. Synonyms helped this example, but a dictionary is not a measure of full language understanding. These results apply to Prism's engine and this fixed example, not to ad placement on ChatGPT, Google or another platform.
Load your landing page or type your existing ad, then try questions your buyers would ask. You can request a PDF of your results after the test. The current PDF downloads in your browser.