Case study
Why the prices in your video ad's captions come out wrong
When captions are written from what a machine heard, a price can appear ten times its real size and a sum meant to stay hidden can sit in plain sight. Our engineer traced it to the transcript, and changed where the on-screen figures come from.
Key takeaways
- Captions generated from a voice transcript spell figures their own way, so on-screen text should come from your script and be checked against it.
- A price written with a currency sign and a thousands comma can come back as two plain numbers, which a cover rule built for one word will miss.
- A dropped decimal point turned one figure into ten times its real size, and nothing in the audio sounded wrong.
- A sum that is already covered still needs to be treated as a sum, or it will never stand out on screen.
- Reading an ad once with the sound off is the quickest way to see a figure that does not match what was said.
A price is the line in a video ad that a viewer reads twice. When the figure on screen is not the figure the presenter just spoke, the ad has put a false offer in front of everyone who watches it, and the people who made it are the least likely to see the mistake, because the voice says the right number. We were asked why the video ad captions in a client's short ads were showing the wrong price, or sums that did not match their script; the faulty ad was reported to us, the highlighting fix and the fix for split and spoken sums went out the same afternoon, and the price cards that night.
What a wrong figure on screen does to an ad
A wrong sum on screen is a false offer, and a sum meant to stay covered can get the ad turned down.
Ad platforms often refuse ads that show earnings figures, which is why some sums in an ad are meant to stay covered. When those sums appear in full instead, the ad can be rejected, and the account that runs it can be restricted. A viewer who reads a figure that differs from the one spoken has also been shown something untrue, whatever the rest of the ad says.
The mistake hides well. The presenter says the right number and the caption beside them says another. Played at full speed with the sound on, the ear wins and the screen goes by unread.
How video ad captions ended up with the wrong price
The script was right; the captions were built from what a machine heard, and the machine wrote the sums its own way.
Our engineer laid the captions of the faulty ad beside its script, word by word, and the script was correct. The captions come from a speech-to-text transcript of the finished voice, and that transcript had rewritten the figures:
- A price the script wrote with a currency sign and a thousands comma came back as two bare numbers with a space between them. The cover rules looked for a sum written as one word, so they found two ordinary numbers and covered nothing. In one finished ad, prices written this way were shown in full.
- A sum with a decimal point and a size word was heard without its decimal point, and only the first part of it was matched. The figure on screen read ten times its real size, with the size word left standing on its own.
- A long sum said in words was split at the and, so only its end was treated as money, and a price said in words lost its first half.
Why some sums never jumped out
Two further causes had nothing to do with the transcript. A sum that was already covered was never highlighted at all, and when a sum shared a caption with another word the ad had emphasised, that other word won.
What we changed
The money rules now read every way the transcript writes a sum. Numbers in a row that together make one amount are treated as one amount. A size word such as thousand or million right after a sum belongs to it. The and inside a number said in words stays part of it. A hyphenated number counts as one number.
With that in place, a cover sticker covers such a sum whole, and every sum of money in the team's own ads jumps out on screen as it is said, ahead of any other emphasised word. A covered sum jumps out as its sticker.
For ads that show prices openly, each price written in digits in the script now gets a card of its own: an old price is struck through, and the offer price is the one a viewer reads last and largest, so the offer is never lost behind a higher figure. The card always shows the script's own words, never the transcript's spelling, never covers the presenter's face, and sums that are not prices get no card.
What the business has now
The change is live on the platform the client's video team uses every day. Sums written in any of the ways the transcript produced them are now covered whole or shown whole, in the script's own figures. The videos the client already had for their own customers render exactly as before, and we checked that they do.
What you can check yourself
- Watch your ad once with the sound off and read every figure on screen against your script. A mismatch shows up quickly this way.
- Look hardest at prices with thousands separators or decimals, and at sums followed by thousand or million. These are where transcripts go wrong.
- If your captions are generated, ask where the figures on screen come from: the script, or the transcript of the voice.
- Check that the price a viewer sees last and largest is the offer, not an old price or a competitor's.
- Watch any ad that hides a sum and confirm the cover sits over the whole figure, currency sign and all.
What to do next
Before your next ad goes out, have someone read the finished version once with the sound off, with your script beside them. If your team works from a template that adds stickers or highlights to sums, ask them to try it with a price written with a currency sign and a comma, with a decimal, and in words, because those are the three shapes that broke here. And when a template changes, keep one of your earlier ads to hand and play it again afterwards, so you can see for yourself that everything your customers already have still looks the same.
Questions people ask
Why do my auto-generated captions show the wrong number when the voice says it right?
Generated captions are usually built from a speech-to-text transcript of the finished audio, and that transcript writes numbers the way it hears them, not the way your script writes them. It can split a price into two separate numbers, drop a decimal point or leave out a currency sign. Your script is not consulted at all, so the voice can be correct while the text beside it is not.
How do I keep a price from appearing in an ad when the platform rejects earnings figures?
The cover needs to find the sum in every shape a transcript produces, not only when it is written as a single word. Two numbers separated by a space, or a figure missing its currency sign, will slip past a rule that expects one written form. After any change, watch the finished ad with the sound off to see that the cover is over the whole figure.
Should prices in my video script be written in digits or in words?
Write them in digits, and treat the script as the only source for anything that appears on screen. A price spoken in words can be split at the and or lose its first half when it is transcribed. A price written in digits gives whoever builds the captions something exact to copy, and someone can then check the finished ad against it.
How can I check an ad's on-screen figures before it goes out?
Play the ad with the sound off, script in hand, and read every figure that appears. Pay closest attention to prices with thousands separators or decimals, and to sums followed by thousand or million. Then confirm that the largest and last price a viewer sees is the one you are actually offering. A few minutes of this catches what the sound hides.