Quality Review and Real-Time Search
Settings for having the AI automatically check quality before publishing, plus fact-checking against current information
Automatic checks before publishing
Enabling "Settings > Content Quality Review" has the AI score each generated article.
| Item | Setting key | Description |
|---|---|---|
| Automatic review | review_enabled | Whether to run a quality check after generation |
| Passing line | review_threshold | Anything scoring below this is treated as having a problem |
| Handling of low scores | review_hold | Whether to hold the article as a draft instead of publishing it |
The check covers whether the article answers the search intent, whether claims are backed up, whether the content is thin, and whether there's duplicated wording.
If your post status is set to "Publish" for fully automatic operation, we strongly recommend enabling this feature and setting it to hold articles that score low.
Revising after writing (Polish)
Enable Polish under "Settings > Generation defaults" (on by default), and right after each section is written, the plugin machine-measures its readability and has the AI rewrite only the sentences that got flagged.
| Setting key | Description |
|---|---|
polish_enabled | Whether to measure readability after writing and fix only the sentences that got flagged |
It checks eight things. Each one is something you can count, which is more reliable than just asking the AI "is this readable?"
| What it checks | Guideline |
|---|---|
| Sentence length | Once a sentence passes 80 characters, split it into two at a conjunctive particle |
| Repeated sentence endings | If 「〜です。〜です。〜です。」("...desu. ...desu. ...desu.") repeats three times, change the third sentence to something like 「〜でしょう」("...deshou," probably) or 「〜かもしれません」("...kamoshiremasen," maybe), or end it on a noun (taigen-dome) |
| Mixed styles | If the polite desu/masu style and the plain da/dearu style are mixed together, bring whichever is used less in line with the more common one |
| Demonstratives | If a lot of sentences pick up with a demonstrative like 「これには」("for this,"), replace it with the actual noun it's referring to |
| AI-sounding phrasing | Things like 「いかがでしたか」(ikaga deshita ka, "how was it?"), 「羅針盤」(rashinban, "compass"), 「〜の実施が重要です」(~ no jisshi ga juyo desu, "implementing ~ is important"), 「地味に効く」(jimi ni kiku, "quietly pays off"), 「解像度を上げる」(kaizodo o ageru, "raise the resolution") or a leading 「重要なのは、」(juyo na no wa, "what matters is") (see "Spotting AI-ness" below) |
| Sentences with many commas | If a sentence has four or more 読点 (Japanese commas), split it or reorder it |
| Spaces between Japanese and Latin text | If a section has three or more spaces like 「Google 検索」, remove them (URLs, code and English sentences are left alone) |
| Decorative "not A but B" | From the second occurrence when a section has three or more. If dropping A doesn't change the meaning, state B directly |
| Chains of の | If の repeats three times in a row, like 「僕の名前の由来の」(boku no namae no yurai no, "the origin of my name's"), restructure the sentence |
| Repeated conjunctions | The second and later occurrences of the same conjunction, as in 「たとえば〜。たとえば〜。」("For example, ... For example, ...") |
| Words you can cut | If 「〜すること」("the act of doing ~") or 「〜という」("called ~") show up often, cut them |
The key point is that it never has the AI rewrite the whole section. Handing over the whole thing would change sentences that didn't need fixing too, and break numbers or links. Instead, only the flagged sentences get passed along, and before the returned section is adopted, the plugin checks that the heading text, link destinations, and the number of tables, bullet lists, and decorations all still match, that the character count is between 80% and 1.25x of the original, and that the number of flagged issues has actually gone down. If even one of those breaks, the original text is used as is.
Only sections with flagged sentences get one extra AI call for the body text. If nothing gets flagged, nothing happens.
Spotting "AI-ness"
Readers don't notice AI writing because of grammar mistakes. These checks apply to Japanese articles: the Japanese is grammatically correct, but readers can still tell from habitual phrasing. So the plugin machine-checks for the following and flags them.
| What it looks for | Examples |
|---|---|
| Stock greetings | 「いかがでしたか」(ikaga deshita ka, "how was it?"), 「参考になれば幸いです」(sanko ni nareba saiwai desu, "I hope this was helpful"), 「ぜひ一度試してみてください」(zehi ichido tameshite mite kudasai, "please give it a try") |
| Empty metaphors | 「羅針盤」(rashinban, "compass"), 「架け橋」(kakehashi, "bridge"), 「車の両輪」(kuruma no ryorin, "two wheels of a cart", an idiom for two equally essential parts), 「ゲームチェンジャー」(gemu chenja, "game changer"), 「土台となる」(dodai to naru, "becomes the foundation") |
| Translation-style phrasing | 「〜の実施が重要です」(~ no jisshi ga juyo desu, "implementing ~ is important"), 「〜を抑制する」(~ o yokusei suru, "to suppress/restrain ~"), 「〜と言えるでしょう」(~ to ieru deshou, "it could be said that ~"), 「〜に他なりません」(~ ni hoka narimasen, "is nothing other than ~") |
| Metaphorical verbs | 「地味に効く」(jimi ni kiku, "quietly pays off"), 「安全側に倒す」(anzen-gawa ni taosu, "tip toward the safe side"), 「時間を溶かす」(jikan o tokasu, "melt time away"), 「1つずつ潰す」(hitotsu zutsu tsubusu, "crush them one by one"), 「静かに壊れる」(shizuka ni kowareru, "silently breaks") |
| Empty buzzwords | 「解像度を上げる」(kaizodo o ageru, "raise the resolution"), 「腹落ち」(haraochi, "click in your gut"), 「肌感」(hadakan, "skin feel"), 「温度感」(ondokan, "temperature"), 「本質的な」(honshitsuteki na, "essential") |
| Overblown words | 「真理」(shinri, "truth"), 「境地」(kyochi, "state of mind"), 「極致」(kyokuchi, "pinnacle"), 「深淵」(shin'en, "abyss") |
| Label openers | A sentence starting with 「重要なのは、」(juyo na no wa, "what matters is"), 「正直に言うと、」(shojiki ni iu to, "honestly speaking") or 「注目すべきは、」(chumoku subeki wa, "what deserves attention is") |
| Decorative symbols | Emoji and long dash characters |
Words used in their literal sense are not flagged (a high-resolution monitor, the calories in rice, a house's foundation, medicine that works). 「結論から言うと」("to give the conclusion first") is not flagged either, since people use it often and it fits the conclusion-first style.
The checks and fixes for metaphorical verbs, buzzwords, label openers, symbols, commas, spaces and "not A but B" are based on yomiyasu (MIT License), a skill for making AI-written Japanese easier to read. Following yomiyasu's rules, the AI is told to keep the claim, what matters most, how firmly it is stated and what each sentence does (explain, evaluate, request), and not to add information that wasn't there.
There's one more check: whether sentence lengths are too uniform.
This one is easy to overlook. Keeping to "one sentence, up to 50 characters" makes an article more readable. But once every sentence comes out at 50 characters, the uniformity itself makes the text read as mechanical. People naturally mix long and short sentences, so that very evenness becomes an AI tell. The plugin measures how much sentence length varies and flags the article when it's too flat. Short articles with fewer than 12 sentences are exempt, since a short piece can end up uniform by chance alone.
The AI is given the same guidance when it generates articles, so this check normally doesn't trigger. It tends to trigger when additional instructions pin the writing style down in detail, or when you've edited the article by hand yourself.
- You can drop 「いかがでしたか」 and lose nothing. For a closing line, write one concrete thing the reader should do next
- Replace metaphors with concrete facts or numbers. Turn metaphorical verbs into everyday words while keeping their nuance
- Instead of opening with 「重要なのは、〜です」("what matters is ~"), start from the point and say it in the predicate, like 「〜が大事です」("~ is important")
- Turn phrases like 「〜の実施が重要です」 into native verb forms, such as 「〜してください」 ("please do ~")
- When the lengths are too uniform, mix in some short sentences. Let your punchier, declarative ones run 10 to 20 characters
Real-time search (fact-checking)
Before writing an article, the AI can actually search the web to confirm information. Choose which search engine to use under "Settings > Real-Time Search."
| Item | Setting key | Description |
|---|---|---|
| Search provider | grounding_provider | Gemini (Google Search) or Grok (Web + X) |
| Grok's search scope | grok_x_search | Whether to also include X posts, when using Grok |
Fact-checking prevents an article from being written as if outdated information were still current. If no key is set, articles are generated without search (you can still create articles, but they'll be less current).
Deep research (Pro)
Normal research gathers two things with search before writing: "search intent and competitor analysis" and "facts about the topic". With deep research on, Sorabun does the following before moving on to the outline.
- The AI reads what was gathered and turns points not yet confirmed, points where sources disagree, and what readers will want to know next into questions that search can answer (up to 5)
- It searches to answer those questions
- It repeats steps 1 and 2 once more (stopping early if nothing is missing)
- Finally, it rechecks every fact with search and rebuilds them into lists of "confirmed facts", "things not to write", and "things that couldn't be confirmed". The article is written with this list as its top-priority basis
It helps most for topics with many numbers, prices, or rules, and for comparison and roundup articles. The sources it checks are also added to the links the article is allowed to use.
| Item | Setting key | Default | Description |
|---|---|---|---|
| Deep research | deep_research | Off | "Settings > Real-Time Search". The default when creating articles |
On the "New Article" screen (including bulk creation) you can turn it on or off per article (for keywords and rewrites). Articles created without going through that screen, such as scheduled posts and the AI Editor-in-Chief, follow the setting.
Each article makes up to 5 more AI calls (3 of them with search), and research takes a few minutes longer. The AI cost of the research stage roughly doubles or triples. If Grok is your search provider, search fees based on the number of sources referenced rise accordingly.
Article quality improves, but this doesn't guarantee that factual errors are eliminated. Keep checking before publishing (quality review and your own review). If an AI call fails during deep research, article creation doesn't stop; it continues with the material from normal research.