We run Liquid AI's d1 decision model as a cheap second worker inside our SEO pipeline. It reads paragraphs and keyword lists, answers yes-or-no questions about them as probabilities, and hands the results to Claude, which does the writing and makes every call. In its first 279 decisions it cost us $0.0021.
Key takeaways
- Liquid's d1 does not write. It scores. That is exactly why we use it.
- Right now it works in two places. One lane lints paragraphs. The other tags law-firm keywords by intent and city.
- 279 decisions on October 27 cost $0.0021 in total, with zero fallbacks to Claude.
- It over-flags short paragraphs, so it points at problems. It does not block a publish.
- A new lane stays in shadow for seven days and goes live only at 90% agreement with Claude.
Why a second model at all
The trigger was a bad number. In October we ran our own client pages through an AI-text detector, and most of the long-form copy came back flagged as machine-written. One pricing page scored 1% human. A comparison page scored 0%. Our own North Carolina service page was in the same shape.
The fix for that is rewriting, and rewriting is Claude's job. But before you can rewrite, you have to find the bad paragraphs, and a 2,500-word page has 70 of them. Sending each one to a frontier model with a seven-point checklist is slow and it is the wrong tool. Most of those questions have one-word answers. Does this paragraph open by restating its heading? Is it hedging? Could it sit unchanged on a competitor's site?
Those are classification questions. So we went looking for a model built to answer them, not to chat about them.
What a decision model is
Liquid AI calls d1 a decision model. You give it a state (a paragraph, a keyword, a JSON row) and a set of typed questions. It reads everything in one pass and returns probabilities instead of text. Three question types cover most of what we need:
- Noul: yes or no, answered as a number between 0 and 1.
- Choice: one label from a set, with a probability for each label.
- Score: a position on an ordered scale.
According to Liquid's launch post for d1, a text decision takes 200 to 300 milliseconds and generates no tokens at all. No tokens means no prose to parse, no malformed JSON, no model deciding to add a friendly sentence before the answer. We get a float back and our code compares it to a threshold.
That shape suits SEO work far better than most people expect. Half of an audit is sorting: which queries are informational, which page should own which keyword, which paragraph is filler.
The two lanes running today
| Lane | What it reads | What it asks | First run | Cost |
|---|---|---|---|---|
| ai-tell-lint | One paragraph at a time from a draft or live page | 7 yes/no questions: restates the heading, stock phrases, stacked hedges, uniform sentence length, no stance, generic, brochure tone | 50 paragraphs in the daily build routine | $0.00038 |
| Intent and city tagging | Candidate keywords for californialawfirmseo.com | Informational or commercial, and whether the phrase is tied to a California city | 44 keywords | $0.00033 |
The lint lane is the one doing the most work. Its questions are written in plain English: "Could this paragraph appear unchanged on a competitor's website, with no specific number, place, customer detail or constraint?" A score of 0.8 or higher on restating, stock phrases, hedging or brochure tone sends that paragraph back to Claude for a rewrite from real inputs: customer questions, account data, job details.
The keyword lane feeds our California law firm SEO sprint, where we need 30 or more targets that are low difficulty, California-specific and not already owned by an existing page. Liquid tags the candidates. Claude reviews the tags against Search Console before anything goes into the plan.
What it actually costs
Every Liquid call goes through one helper script that logs to a single spend ledger. Here is the whole of it so far:
| Date | Runs | Decisions | Input tokens | Claude fallbacks | Spend |
|---|---|---|---|---|---|
| 2026-10-27 | 8 | 279 | 105,297 | 0 | $0.0021 |
A fifth of a cent. The monthly ceiling we set is $10, and at this rate the ceiling is irrelevant. The real cost is the engineering time to write good questions, which is the same work you would do writing a rubric for a human editor.
The fallback column matters more than the spend. If Liquid returns an error or a result that fails our schema check, the item goes to Claude in one structured call so a routine never stalls. On day one, nothing fell through.
Where Liquid was wrong
The interesting part. We ran the lint lane on three pages before and after rewriting them, and compared it with two other signals: our own tell score (a local script that measures rhythm, stock phrases, hedges and question headings) and the Sapling detector.
| Page | Tell score before → after | Detector human share before → after | Liquid: paragraphs with any flag ≥ 0.5 |
|---|---|---|---|
| Catering pricing page | 75.5 → 94.4 | 1% → 100% | 100% → 76% |
| Office water comparison | 65.2 → 93.2 | 0% → 100% | 74% → 95% |
| Our NC service page | 76.2 → 86.8 | not run → 100% | 83% → 85% |
Look at the office water row. The detector swung from certain-AI to certain-human, and Liquid's flag rate went up. We traced it to the rewrite itself: good copy uses short paragraphs, and Liquid hates them. Four words of transition carry no number and no place, so it gets tagged generic. Technically correct. Useless as an instruction.
So we split the seven questions in two. Four of them (restating, stock phrases, hedging, brochure tone) were right often enough to trigger a rewrite at 0.8. The other three now only highlight a paragraph for a person to glance at. We'll know by November 3 whether that split holds.
Where Liquid is allowed to run
A model this cheap gets used everywhere unless something stops it. Ours sits behind a 474-line Python helper, and the helper is the only way in. No routine calls Liquid's API on its own, which is how we can say the ledger above is complete rather than roughly right.
What Liquid may touch is narrow. It scores and it tags. Article prose, negative keywords, ad budgets and publishing are off its list entirely; Claude makes those calls, and a person signs off before any existing page on a client site changes.
New lanes don't start live. For the first week Claude decides every item exactly as before, and Liquid's answers sit next to Claude's in a log. If they match on nine items in ten, the lane gets promoted. If a promoted lane later slips under that line for a week, the script quietly hands the work back to Claude. Nobody has to catch it on a Monday call.
"Our focus on the quality of content, rather than how content is produced, is a useful guide."
Google Search Central, Google Search's guidance about AI-generated content (February 8, 2023)
Why care about detector scores at all, if Google judges the content and not the tool? Because the score was never the target. On the three pages we tested, the 0% and 1% scores landed on copy that could have described any business in any city, and that copy loses with readers and with AI answer engines long before it loses with a detector.
If you want to try this
Start with the questions, not the model. Our lint prompt started with nine flag codes. Two of them, counting em-dashes and spotting empty lists of three, our local tell script already measures with plain pattern matching, so the version Liquid scores asks seven.
Pick tasks where the answer is a category or a probability. Our next lane is the 28-day check on title and meta changes: Search Console says whether clicks went up, held or fell, and Liquid tags each change win, flat or loss so Claude only reads the losers. If you catch yourself wanting the model to explain its answer, the task belongs to the bigger model.
Keep the judge. The value of d1 is that it lets Claude spend its time on the 20 paragraphs that need rewriting instead of reading all 70. It is a filter in front of judgment, not a replacement for it. If you want to see how that fits the rest of our AI SEO work and our generative engine optimization for clients, those pages walk through the full stack. For keyword data specifically, our note on AI keyword research tools covers why a model's volume guess is never the source of truth.
FAQ
What is Liquid AI's d1 model?
d1 is a hosted decision model from Liquid AI. It takes text or images plus typed questions and returns probabilities (yes/no, a choice among labels, or a score on a scale) without generating any text.
Does Liquid AI write content for ByteZero clients?
No. Claude writes, and real inputs from the client supply the specifics. Liquid only classifies and scores, and its output is reviewed before anything changes on a page.
How much does it cost to run?
Our first 279 decisions cost $0.0021 in total. We cap Liquid at $10 a month, and every call is logged with the routine and task that made it.
Is AI-detector score a Google ranking factor?
No. Google says it rewards helpful, high-quality content however it is made. We use detector and lint scores as a proxy for generic copy, which readers and AI answer engines both skip.
Can a decision model replace an SEO editor?
It replaces the part of editing that is sorting. Deciding what to say, what to cut and what the client can stand behind stays with Claude and with us.
Can you run this on our site?
Yes. It is part of our search engine optimization work. A free AI visibility audit is the usual starting point.
