Anyone who has tried to push a single AI prompt past the 1,500-word mark knows the failure mode: the model starts strong, drifts around the 800-word mark, and limps to the finish line with thin, repetitive paragraphs. That’s not the model being “lazy” — it’s a structural problem. Large language models pay a quality tax when generating long outputs in a single pass, and most articles end up shorter than the requested length not because of bad prompting but because of the architecture of one-shot generation itself. AI Content Generator solves this with a two-stage pipeline: chunk generation followed by global review. This tutorial explains how it works and how to dial it in for your campaigns.
By the end of the guide you’ll know exactly which length target band to pick, how to size the chunks, what the global review pass actually does, the per-section minimum word count that prevents thin H2s, and the trade-offs between generation time and final quality. The settings are all per-campaign, so different topical clusters can have completely different length strategies.
The pay-off is concrete. On a 2,000-word target band, AI Content Generator with chunk + global review produces articles averaging 2,050 words versus 1,200 for a single-pass prompt with the same model — and with one-quarter the repetition rate. The trade-off is generation time (around 35–45 seconds instead of 15–20) and a higher token cost, but for long-form SEO content where retention and ranking depend on depth, the trade-off is almost always worth it.
Why single-pass long-form fails
A language model generating a long article in one pass has two structural problems. First, the further the model gets from the start of the prompt, the weaker its attention to the original instructions becomes — by the time it’s writing section 6 it has forgotten the tone instruction from section 1. Second, models have an implicit “expected output length” that pulls them toward 800–1,200 words for any natural prompt, even when you explicitly ask for more. The combination produces articles that are not only shorter than requested but also visibly drift in tone and structure.
Chunk generation breaks the article into 2–3 H2 sections per call, so the model writes each chunk with full attention to the brief and a much shorter expected output. Global review then takes the stitched-together draft and runs one more pass dedicated entirely to balance, deduplication and transitions. Two passes, four to five API calls, one polished article. That’s the whole architecture.
Step 1 — Open the SEO brief (content) panel
Inside a campaign editor, scroll to the panel titled SEO brief (content). The length controls all live in the sub-section Article length & generation strategy. Every setting here applies to every article this campaign produces from now on — change it once and it propagates.
Step 2 — Pick the right target length band
The Target length dropdown offers five bands: Short news (300–600), Medium (700–1,200), Long-form (1,500–2,000), In-depth (1,800–2,200), Pillar (2,500+). Bands are bands, not exact numbers: the plugin keeps the final article inside the range with high reliability but doesn’t try to hit a specific count.
Pick the band that matches the editorial intent of the campaign, not the keyword competition. A how-to tutorial in a competitive niche still wants Long-form; a quick news brief on a hot topic stays at Short news even if the keyword is competitive. Forcing 2,500 words on a topic that doesn’t deserve them produces padding, and Google penalises padding faster than it rewards length.
Step 3 — Configure chunk size
The Chunk size dropdown controls how many H2 sections the plugin asks the model to write per API call. The default is 2 H2 sections / call — that’s the sweet spot for most models and most lengths. Smaller chunks (1 section per call) produce higher-quality individual sections but slow down generation and cost more in tokens. Larger chunks (3–4 sections per call) speed things up but reintroduce the drift problem.
Two practical recommendations. First, leave the default at 2 unless you have a specific reason to change it — three years of internal testing across many models points consistently to 2 as the optimum. Second, if you switch to a smaller model (a “Flash” or “Mini” variant), dropping the chunk size to 1 can recover meaningful quality at the cost of speed. Larger models almost never benefit from chunk size 1.
Step 4 — Enable the global review pass
The Global review toggle is the single most important quality setting in the entire campaign editor. When enabled, after the chunks are stitched together, AI Content Generator makes one final API call that asks the model to read the complete article and do four specific things: balance the lengths of H2 sections so no section is more than 30% shorter or longer than the average; deduplicate sentences and phrases that the chunks accidentally repeated; smooth the transitions between sections; verify that the tone and the structural instructions from the brief are respected throughout.
Disabling global review saves about 10 seconds per article and a few cents per generation but produces visibly weaker output. Leave it on. The only legitimate reason to disable it is high-volume short-form publishing (200-word news briefs at hundreds per day) where the review pass doesn’t have enough material to improve.
Step 5 — Set the per-section minimum word count
The Min words per H2 section field is a hard guarantee: no H2 section will ever be shorter than this number. For a Long-form target (1,500–2,000 words across 7–8 sections), 180–220 words per section is right. For In-depth (1,800–2,200 across 8 sections), 220 is comfortable. For Pillar (2,500+), bump to 280.
The minimum interacts with the number of H2 sections you configure. If you set 8 sections and 220 words minimum, the floor is 1,760 words — anything below that and the global review re-asks the model to expand the weakest sections. The plugin reports in the Logs whenever this expansion is triggered, so you can see which articles needed extra work.
Step 6 — Aim for the upper bound of the target band
By default models drift toward the lower bound of any length band you give them. The Aim for dropdown counters this: setting it to Upper bound of target band changes the implicit prompt the plugin sends from “stay between 1,500 and 2,000 words” to “aim for ~2,000 words while staying under 2,000”. Same band, completely different average output length.
The upgrade is essentially free: same number of chunks, same global review, same cost per article — but the average length climbs by 200–400 words. For SEO-oriented content where the band exists precisely to maximise depth without padding, “Upper bound” is the right default.
Step 7 — Use Extra AI instructions to reinforce length discipline
Settings cover the structural side of length control. The behavioural side lives in the Extra AI instructions (campaign) textarea. A few short instructions go a long way:
- “Each H2 section must have at least 2 substantial paragraphs, no shorter than 4 sentences each.”
- “Do not use one-sentence paragraphs. Combine related thoughts into proper paragraphs.”
- “When you reach a natural stopping point in a section but the section is still short, expand with a concrete example rather than abstract restatement.”
- “Never start an H2 section with a one-line summary; jump directly into substance.”
These four lines change the texture of generated articles overnight. The global review pass uses the same instructions on the second pass, so they reinforce themselves.
Step 8 — Inspect a generation in the Logs panel
After running a campaign with these settings, open AI Content Generator → Logs. For each article you’ll see a per-stage breakdown: how long each chunk took, how many tokens it cost, whether the global review triggered a section expansion, the final word count. This is the dashboard to use when you’re tuning a campaign — staring at the logs for the first 10 articles tells you exactly which settings to nudge.
The most common things you’ll spot. Articles consistently landing at the lower bound of the band: enable “Aim for upper bound”. Sections of wildly different lengths: lower the chunk size to 1, raise the global review temperature slightly. Repetitions in the final text despite review enabled: increase the frequency penalty in the AI provider settings.
Troubleshooting
- Articles are shorter than the band — set “Aim for: upper bound” and raise the minimum words per section.
- Global review never triggers expansion — your minimum-per-section is too low. Raise it.
- Generation takes more than 60 seconds — drop chunk size from 1 back to 2, or switch to a faster model for the chunks (keep the smart model for review).
- The final article has dangling references (“as mentioned above” when nothing was mentioned) — global review couldn’t reconcile the chunks. Lower the chunk size to 1 so each chunk is more self-contained.
- Cost per article is too high for the band — switch the chunk model to a cheaper variant (Mini/Flash/Haiku) and keep the heavyweight model only for the global review pass.
You’re done
That’s the full configuration: pick the right target length band, leave chunk size at 2, enable global review, set per-section minimums, aim for the upper bound, reinforce the brief with extra AI instructions, then watch the Logs panel for the first 10 generations to fine-tune.
From here, every article AI Content Generator produces in this campaign lands inside the target band with balanced sections and almost no repetition. The fifteen extra seconds per generation are the price of the quality jump, and the price drops to essentially nothing once you stop having to manually expand thin drafts. Long-form SEO content stops being the slowest part of your editorial pipeline and starts being the most reliable one.