When content costs nothing
Price a generated post with its review, work out what a flood of posts does to visits per page, and decide where generation earns its keep.
On this page
A post that cost $400 to write now costs under $1 to generate and $50 to review. The saving is real, and every competitor gets it on the same day. In a channel whose value came from scarcity, a price cut available to everyone mostly moves the value somewhere else; the arithmetic below finds where, and which generated content still pays for its review.
The cost line, with the review in it
The mistake comes first: "500 posts a month for under $500." That prices the model and nothing else. The arithmetic a marketer carries puts a generated post at under $1 of model cost plus a $50 review, 30 minutes of a $100 hour, so $51, against $400 for one written in four hours.
| Volume | Written at $400 | Generated and reviewed at $51 | Review hours |
|---|---|---|---|
| 8 posts a month | $3,200 | $408 | 4 |
| 100 posts a month | $40,000 | $5,100 | 50 |
| 500 posts a month | $200,000 | $25,500 | 250 |
Production cost falls by 400 ÷ 51 ≈ 8 times, not 400, because the review is the floor; how to cost the model call is in where the money goes, and at the illustrative $3 and $15 per million input and output tokens in the four numbers a 1,500-word post from a 3,000-token brief costs about $0.04. At 500 posts a month the review is 250 hours against a working month of 176, about 1.4 full-time reviewers. Skipping it saves $25,000 and removes the one check that stops an invented statistic going out under the company's name.
A channel whose value was scarcity
Search and content paid because demand for an answer is roughly fixed while good answers were few. Take a topic whose readers produce 40,000 visits a month, and assume 200 comparable pages compete for them: 200 visits a page. Assume, as content and SEO does, that a blog visit is worth a third of a $2 paid search click, $0.67. A page returns about $134 a month, and a $400 post pays for itself in three months.
Now let every publisher generate. Five competitors at 500 posts a month add 2,500 pages a month: 5,200 pages after two months, 30,200 after a year. Demand does not grow with supply, so the same 40,000 visits spread over all of them.
The average page falls to 40,000 ÷ 30,200 ≈ 1.3 visits a month, worth about $0.88. A $51 post needs $51 ÷ 12 ÷ $0.67 ≈ 6.3 visits a month to pay back within a year; the average generated page gets a fifth of that.
The average hides the shape, and the shape makes it worse. The first few results take most clicks and the second page almost none, so the flood makes the typical page worth nothing and leaves the value with the few pages the engine still puts first. The question for a generated post is not whether it is cheap but what will put it among those few, when every competitor can generate the same post for the same $51. Answer engines tighten this further, since a question answered on the results page sends no visit at all; search after the answer engine prices being cited.
Where the value moves
When production is cheap for everyone, the scarce inputs are the ones a generator cannot produce for a competitor.
Distribution. An owned audience does not compete for the topic's 40,000 visits. The course's email send reaches 20,000 users and produces 40 conversions for $20 of sending and an hour of writing; the same post published into the flooded topic starts at 1.3 visits a month. Multiplying output without multiplying distribution adds pages to a field that has grown 150 times.
Proprietary data. The clinic scheduler from segmentation has 400 clinics' booking records, enough to publish "a two-reminder sequence cut no-shows from 12% to 5%". A competitor's model can write about no-shows; it cannot produce that figure. Answer engines cited pages carrying a statistic more often on the first peer-reviewed benchmark and not on the second, both weighed in search after the answer engine, so the case for this one rests on the plainer fact that no competitor can publish it.
Trust. A named author, claims that held, reviews from real customers. It accrues slowly and one careless post spends it, which is the case for the $50 review on everything that ships.
What the evidence says, and where it is thin
The search engines' own guidance is the firmest ground: how a page was produced matters less than why. Google Search's guidance about AI-generated content (opens in a new tab) commits to rewarding high-quality content "however it is produced", and calls automation used mainly to manipulate rankings a violation. Google Search's guidance on using generative AI content on your website (opens in a new tab) adds that using generative tools to "generate many pages without adding value for users" may violate its spam policy on scaled content abuse. A 500-post plan whose pages exist to collect search traffic is the case that policy describes.
On demand, Large language models reduce public knowledge sharing on online Q&A platforms (opens in a new tab) (del Rio-Chanona, Laurentsyeva and Wachs, PNAS Nexus, 2024) estimates that activity on a large programming Q&A site fell 25% within six months of the first widely used chat assistant's release, compared with similar sites where access was restricted: the fixed demand above, shrinking, as answers from a model replace pages.
On whether generated posts rank and convert as well as written ones, the evidence is thin: the comparisons we found come mostly from companies selling generation or search tools, rarely describe a control, and none is peer-reviewed. Measure it on your own site: generate posts for a random half of a topic list, write the other half, and compare visits and sign-ups per post at 90 days.
Where generated content is worth producing
Generate where the value does not depend on being one of few; write where it does.
- Worth generating: pages with specific demand and thin competition, built from the company's own facts: one page per integration from its documentation, help answers from its support tickets, lifecycle emails to its own list. A check against the source catches most errors.
- Worth generating as a draft: the first version of a piece built on proprietary data, which a person then checks, cuts and signs.
- Not worth generating: general posts every competitor can also generate, such as "ten ways to reduce no-shows", at $51 for 1.3 visits a month.
In practice
When a founder or an agency proposes volume, ask for the cost with review counted, the visits per post assumed, and what the posts carry that a competitor's model cannot: "500 posts at $51 is $25,500 a month and 250 review hours, and each post needs about 6 visits a month to pay back in a year. Which topics are thin enough, and which of our own numbers is in the posts?" To the team: write the few pieces built on the company's numbers, generate the pages its sources can check, and spend the saving on distribution.
The mistake that costs money is pricing the plan at model cost and valuing it at the old visits per post. 500 posts at 200 visits each and $0.67 a visit looks like $67,000 a month of traffic for $500. With the review counted and the topic flooded, each month's batch costs $25,500 and earns about 500 × 1.3 × $0.67 ≈ $435 a month.
Check yourself
1. An agency offers 60 generated, reviewed posts a month for $3,200, the price of your 8 written ones. Your topic's pages average 200 visits a month today, and three competitors have started publishing at volume. Which do you buy?
Not the 60 on volume. If the topic floods toward the worked example's 1.3 visits a page, 60 posts earn about 60 × 1.3 × $0.67 ≈ $52 a month for $3,200. Keep the 8 written posts only if each carries your own data; buy generated pages only where your sources can check them and the agency can name each page's demand.
2. Your founder wants 500 posts a month. Estimate the monthly cost and the visits each post must average to pay back within a year, then decide.
500 × $51 = $25,500 a month, almost all of it 250 review hours. Each post needs $51 ÷ 12 ÷ $0.67 ≈ 6.3 visits a month, so a year's 6,000 posts need about 37,800 visits a month, nearly all of a 40,000-visit topic that competitors also want. Decline the volume; fund the 20 or 30 pages your own data can support.
3. Three proposals: 300 integration pages generated from your API documentation, 300 general guides on industry topics, and a monthly benchmark report from your customers' usage data. Which do you generate, which do you write, and which do you drop?
Generate the integration pages: specific demand, thin competition, and the documentation to check them against. Write the benchmark report, with generation for drafts at most; its value is the data, which no competitor's model can produce. Drop the general guides: 1.3 visits a month against $51.