A PM once sent me a lovely rate. Then she sent the grid. The grid clawed back about a third of it before I'd translated a single new word.
This is the part of the business nobody argues about hard enough. We obsess over the headline per-word number — the one we put on our profile, the one we fight for in the "can you sharpen your rate" email. Meanwhile the discount schedule attached to it quietly decides what you actually get paid. For a Turkmen project full of repetition, the grid matters more than the base rate. For a 40,000-word technical manual with a half-built TM, it's the whole game.
What the standard grid really costs you
You know the shape of it. The Trados/memoQ default most agencies push looks roughly like this: 101% and 100% matches at 25% of rate, 95–99% at 30%, 85–94% at 60%, 75–84% at 60%, no-match at 100%. Repetitions bundled in with the high-fuzzy band.
Fine in theory. The TM did some work, you do less, you charge less. I don't object to the principle. I object to the numbers, because they assume the TM is clean and the context is stable — and neither is true often enough to justify the scale.
Take 100% matches at 25%. A 100% match means the source segment is identical to one already in the memory. It does not mean the translation fits. Different sentence, same words, different gender agreement, different register because now it's a button label instead of body text. In Turkmen, a verb form that was right in a past-tense procedure is wrong in an imperative UI string. I still have to read it, decide, sometimes retype it. And I'm being paid a quarter for the reading, which is the part that takes the actual time.
Then there's the 95–99% band at 30%. A 99% match can be a changed number, or it can be a flipped negation. "Do not connect" versus "connect." Same fuzzy score. One is a safety instruction. Catching the difference is worth more than 30%, not less.
The fuzzy band is where the trap lives
The 85–94% range at 60% is the one that quietly bleeds you. A high fuzzy looks like a gift and often eats more minutes than a blank segment. With a no-match I read the source and write Turkmen. With an 86% match I read the source, read the suggestion, find what changed, decide whether the old structure still holds after the change, and rebuild the sentence around it. Turkmen is agglutinative — you change one element and the suffixes downstream move with it. Editing a near-match isn't lighter work than writing fresh. Sometimes it's heavier, because you're fighting someone else's sentence instead of building your own.
So my position: I'll accept a grid, but I push the 95–99 band up, and I treat 85–94 as full rate unless I've seen the TM and trust it. If the memory is a client's own, maintained, consistent — I'll bend. If it's a "we aggregated this from three vendors over five years" TM, the matches are landmines and I price them like new work, because that's what they are.
Repetitions I'll usually give, with one condition: that I actually get to propagate them and they're genuinely identical in context. Autopropagate across a spreadsheet where row 47 means something different from row 12 is not a repetition, it's a bet.
How to argue it without losing the client
The move isn't to refuse the grid. It's to make the grid a conversation about the TM, which flips the burden. When a PM sends a scale, I ask three things. Who owns this memory and is it maintained? What's the repetition and high-fuzzy percentage in the actual analysis? And does my price include running my own QA over the propagated matches, or are those delivered as-is and off my liability?
That last one matters. If I'm paid 25% for 100% matches, I'm not promising those segments are correct — I'm promising I didn't break them. Agencies want full accountability at a quarter price. You can't have both, and saying so plainly, once, usually ends the squeeze. Not because you've been clever, but because you've named the thing they were hoping you wouldn't.
The other quiet tactic: price by the hour for anything with a messy TM, and let the analysis justify it. The word grid is a proxy for effort. When the proxy lies — and on heavy-fuzzy, dirty-memory jobs it lies a lot — go to the thing it's a proxy for. Track the hour. I've written about that before and I'll keep saying it, because the grid is the single biggest place the word-count fiction costs you.
I've worked with agencies that have thoughtful grids. Lower top-band discounts, full rate from 84% down, repetitions capped at a sane percentage, context matches flagged separately. Those agencies get my faster yeses and my better deadlines. The ones who send the default Trados scale and call it "industry standard" get a counter-grid back the same hour.
Standard isn't fair. It's just what loaded by default. Change the default.