Content Science: What to Post, in What Form, How Often, and How to Know

The beginner course on the Journalism site teaches the anatomy of a post and a week of posts from one article (anatomy, one article, a week of posts). This lesson asks a harder question of each piece of advice: who showed it, on how many people, with what design, and how large was the effect. The answer changes what you do. A randomized result with a 20 percent effect is worth building a routine around. A vendor average across 70 million brand posts is worth a test on your own page. A tip from a conference stage is worth nothing until it survives one of the first two.

Reading the evidence before using it

Content advice rests on four kinds of evidence, and they are not equal.

Kind	Example	What it can show	What it cannot
Randomized experiment	Upworthy's 22,743 headline tests; Sahni's email field experiments with millions of recipients	Cause: changing X changed Y, with an effect size	Whether your audience behaves like theirs
Meta-analysis of experiments	Braddock and Dillard on narratives (74 studies); Walter et al. on humor (89 studies)	Average effect and its consistency across settings	Which setting you are in; many included studies are students in labs
Platform statements	Instagram's ranking explainer; TikTok's "How TikTok recommends"; Meta's engagement bait policy	What the system is built to reward or punish	Magnitudes; the statements are deliberately vague and change
Vendor benchmarks	Socialinsider's 70 million posts; Rival IQ's 2,100 companies; Sprout's 307,000 profiles	What brands did and how it performed, on average, in that sample	Cause. Brands choose formats for reasons tied to the content, so format and content are confounded

The order matters when advice conflicts. If a benchmark says carousels win and your own two-month test says single images win on your page, your test wins, if it was large enough to mean anything. The last section is about making it large enough.

Format: what the benchmarks show and what they do not

Socialinsider's 2026 benchmark edition covers 70 million posts from January 2024 through December 2025 and defines engagement rate per follower differently on each platform: likes and comments on Instagram; reactions, comments and shares on Facebook; likes and retweets on X; likes, comments, shares and saves on TikTok (Socialinsider). Its 2025 platform averages were TikTok 3.70 percent (the report's table; its text says 3.73), Instagram 0.48 percent, Facebook 0.15 percent and X 0.12 percent. Its Instagram report, from 35 million posts across 447,613 pages in 2025, breaks format out (Socialinsider Instagram):

Instagram format, 2025	Average engagement per follower	Average posts per month per brand	Change from 2024
Carousels	0.55 percent	5	Flat
Reels	0.52 percent	8	Up 0.02 points
Single images	0.37 percent	7	Down 0.08 points
All formats	0.48 percent	20	Down 24 percent year over year

Rival IQ's 2025 report, 150 companies in each of 14 industries selected at random from a database of more than 200,000, with engagement rate defined as interactions over followers, adds the nonprofit detail: nonprofits had above-median engagement on every channel; on Facebook, photos earned the top engagement rate for nonprofits, which posted that type three times as often as any other (Rival IQ's wording is that nonprofits "shared this post type 3x more often," meaning they published it, not that audiences reshared it); on Instagram, carousels led for nonprofits; on X, videos earned top engagement while photo and link posts were more common (Rival IQ, 2025). Across platforms, Rival IQ recorded engagement falling year over year: Facebook down 36 percent, Instagram down 16 percent, TikTok down 34 percent, X down 48 percent.

Three cautions before you reorganize your calendar around that table. First, these are observational: brands post carousels when they have several things to show and reels when they have motion, so format and content travel together, and a benchmark cannot separate them. Second, the Instagram definition excludes shares and saves, which Instagram's own ranking explainer lists as the signals it most watches for Reels and Explore (Instagram, 2023). A format that earns shares but few comments is undercounted. Third, the sample is brands with marketing budgets; an advocacy page with 3,000 followers in one county is in a different distribution, and Socialinsider's own growth data (in lesson 06) shows small accounts behave differently from large ones.

What the table does support, firmly enough to act on, is that a single image with a sentence of text is the weakest common format on Instagram, and that a carousel of several images, or a short video, does better on average by about half again. For an advocacy page, the carousel is a natural fit: page one is the claim, pages two through five are the document pages that prove it, the last page is the ask. That is a receipt, not a meme, and it uses the format the data favors. Where the message is a person or a place, a reel. Where it is a number from a budget, a photo of the budget page with the line circled, which on Facebook is the format with the top engagement rate in the nonprofit sample.

Hooks and the first seconds: a measurement rule, not a law

The "three-second rule" for video comes from Facebook's accounting, not from psychology. As Digiday reported in 2016, Facebook counts a video view at three seconds, and advertisers were charged only for three- or ten-second views; publishers told the outlet that as much as 85 percent of their Facebook video views happened with the sound off (Digiday, May 2016). That is old, publisher-reported data, and it describes an auto-playing feed where a view was cheap. The lesson that survived is structural: feeds auto-play without sound, so a video that opens on a logo, a title card or a person saying "Hi everyone" has spent its free seconds on nothing.

What the platforms actually measure now:

• YouTube's audience retention report shows, second by second, what share of viewers are still watching. Its intro metric reports "what percentage of your audience still watched your video after the first 30 seconds," and YouTube's own advice is to align thumbnail and title with the content and to move compelling material earlier, since audiences shrink over a video's length (YouTube Help). The dips and spikes in that graph are the only real hook data most pages will ever have.
• TikTok says strong signals outweigh weak ones, and its example of a strong signal is "whether a user finishes watching a longer video from beginning to end." Follower count is not a direct factor (TikTok, 2020). Completion, not the opening alone, is the prize.
• Instagram lists, among the predictions it makes for Reels, how likely you are to reshare a reel, "watch a reel all the way through," like it, and go to the audio page (Instagram, 2023).
• Facebook says recommendations favor videos with engagement beyond plays and watch time, including reactions, comments and shares, that reels are "often 90 seconds or less," and that unedited long content reposted from elsewhere is "unlikely to perform well" (Meta for Creators, 2024).

So the opening has two jobs, and they are different. It has to stop the swipe in the first second or two (that is the Facebook three-second economy), and it has to make a promise the rest of the video keeps, because completion and reshares are what the systems predict. A hook that oversells produces a sharp retention cliff at second four, which YouTube will show you as a dip, and a video that is swiped away at second four teaches the recommender to stop showing it. The reel craft is in the Journalism site's reels course; the evidence base here says: open on the thing itself (the document, the place, the sentence of the claim in text on screen), promise one specific payoff, and deliver it before the end.

Captions and accessibility

Captions are a standards matter before they are a growth tactic. WCAG 2.2, the W3C recommendation updated December 12, 2024, requires at Level A that "captions are provided for all prerecorded audio content in synchronized media" (1.2.2) and that "all non-text content that is presented to the user has a text alternative that serves the equivalent purpose" (1.1.1), and at Level AA a text contrast ratio of at least 4.5:1 (1.4.3) (W3C). A government body's page is often bound to those criteria by policy; an advocacy page is not, and should meet them anyway, because the population an advocacy page is trying to reach includes the people who need them, and because the silent-feed finding means captions are how most viewers get the words.

The commonly repeated figure that captioned video earns more watch time comes from a 2016 Facebook statement we could not retrieve this session, so it is not cited here. The publisher figures in the Digiday piece are the verifiable evidence that sound-off viewing was the norm in auto-playing feeds, and the platforms' own caption tools since then are the practical response. The reels course covers how to burn captions in and keep them above each app's bottom overlay (burned-in captions).

Narrative: the transportation evidence

Green and Brock's 2000 studies in the Journal of Personality and Social Psychology defined transportation as absorption into a story, involving imagery, affect and attentional focus. More transported readers held beliefs more consistent with the story, rated its protagonists more favorably, and noticed fewer false notes; when the researchers reduced transportation with processing instructions, story-consistent beliefs fell. Labeling a story as fact or fiction generally did not change transportation or the resulting beliefs (Green and Brock, 2000). That last finding is the one advocacy pages should sit with: readers absorbed in a story update their beliefs whether or not it is true, which is why a fabricated anecdote is both effective and indefensible.

Braddock and Dillard pooled the experimental literature in 2016 (Communication Monographs 83(4)). Across 74 studies selected from 6,427 screened, narratives compared with control messages produced effects of about r = 0.17 on beliefs (37 studies, 7,376 participants) and about r = 0.23 on behavior (5 studies, 978 participants), with the full set of outcomes ranging from 0.17 to 0.23 (figures as summarized by The Decision Lab; the original is paywalled and we could not read it directly this session). On the usual scale, 0.1 is small and 0.3 medium. That is the honest size of the story effect: real, reliable, and modest.

For a page, the application is to lead with a person's account when the issue has one, and to let the account carry the specifics the document proves. A post that opens with a rancher's own words about a canal and then shows the permit page that confirms them uses both effects, transportation and evidence. The limit is in lesson 09: the persuasion literature on political contact finds effects near zero in general-election settings, so the story's job is mobilizing people who already lean your way and framing the issue for people who have not decided, not converting opponents.

Humor

Walter, Cody, Xu and Murphy's 2018 meta-analysis of 89 studies found humor has "a weak and significant effect on persuasion (r = .13)": a moderate effect on knowledge (r = .23), weak effects on attitudes (r = .12) and behavioral intent (r = .09), an inverted-U relationship with intensity (some humor helps, a lot stops helping), related humor working better for highly involved audiences, and no support for a sleeper effect in which humor pays off later (Human Communication Research 44(4)). So a dry line that is about the subject (the related kind) helps people remember the fact, does little to change their minds, and stops working when the post becomes a comedy routine. A page that explains a utility rate case can afford one joke about the 400-page filing. It cannot afford to be funny instead of specific.

Technique	Evidence	Effect size	Sample	Practical reading
Narrative versus plain message	Braddock and Dillard 2016 meta-analysis	r = 0.17 beliefs; r = 0.23 behavior	74 studies; 7,376 participants for beliefs	Lead with a real account; pair it with the document
Humor	Walter et al. 2018 meta-analysis	r = 0.13 overall; 0.23 knowledge; 0.12 attitudes; 0.09 intent	89 studies	One related joke aids recall; more does not
Negative words in headlines	Robertson et al. 2023, Upworthy RCTs	About +2.3 percent CTR per negative word; about -1.0 percent per positive word	22,743 RCTs; about 105,000 headlines; 5.7 million clicks	Small per word; the effect compounds, which is how clickbait works
High-arousal emotion	Berger and Milkman 2012, NYT most-emailed plus experiments	Awe, anger, anxiety raise sharing; sadness lowers it	Three months of NYT articles plus lab experiments	Covered in lesson 01; the page's choice is which high-arousal emotion
Moral-emotional words	Brady et al. 2017, Twitter	About +20 percent diffusion per word, mostly within group	563,312 tweets	Travels among your own; rarely crosses to the other side
Recipient's name in email subject	Sahni et al. 2018, field experiments	Opens +20 percent (9.05 to 10.80); leads +31 percent; unsubscribes -17 percent	Millions of recipients, three firms	The one email tactic with a large randomized result; see lesson 07
Public token support (petition, like) before a real ask	Kristofferson, White and Peloza 2014	Private token support produced more later help than public	Study count not given in the abstract	A public like can satisfy the urge to help; ask for the real action soon after

Visual evidence and documents

Advocacy pages have one structural advantage over brands in the benchmark samples: they have receipts. A budget line, a permit, an email obtained through a records request, a photograph of the thing itself. The Rival IQ nonprofit finding that Facebook photos earned the top engagement rate of any post type, and that nonprofits posted them three times as often as anything else, is consistent with what anyone who has posted a marked-up document knows: people react to proof. Rival IQ measures reactions, comments and shares together against followers, so it does not isolate the share count; whether a document photo earns more shares than other formats on your own page is a test to run with the method at the end of this lesson. Berger and Milkman's finding that surprising and practically useful content is more viral, independent of emotion, points the same way (JMR 2012). A document that shows a number nobody knew is both. The craft is in the how-to above and in the Journalism site's lesson on rebutting a false claim with a screenshot rather than a quote-post (rebutting without spreading).

Calls to action, questions and comment-driven reach

Comments and shares are ranking inputs. Instagram's explainer lists, for Feed, how quickly people like, comment, share and save a post as the second most important signal after your own activity, and for Explore it says those popularity signals matter "much more" than in Feed (Instagram, 2023). Facebook says its recommendations favor videos with reactions, comments and shares (Meta for Creators, 2024). The temptation that follows is obvious and Meta has a rule against it.

Facebook's engagement bait policy, announced December 18, 2017, demotes posts that push people to like, share, comment or tag in order to boost reach (its example: "LIKE this if you're an Aries!"), applies stricter Page-level demotions to Pages that use the tactic systematically, and was later extended to comments and to videos with bait in the audio. Posts asking for help, advice or recommendations are exempt; the announcement's examples were circulating a missing child report, raising money for a cause and asking for travel tips (Meta Newsroom). TikTok's integrity guidelines likewise make "like-for-like" and similar engagement tricks ineligible for the For You feed (see lesson 04).

The exemption is the whole strategy. A real question ("Has anyone read the full rate filing? Which section covers the fixed charge?") earns comments and is excepted. A request for help ("We need someone at Thursday's hearing who can speak to the 2019 flood") earns comments and is excepted. "Comment YES if you agree" is bait, and a Page that does it weekly is demoted as a Page. The line is whether you want the answer.

One more finding complicates the easy ask. Kristofferson, White and Peloza showed that an initial token display of support, "such as wearing a ribbon, signing a petition, or joining a Facebook group," led to more subsequent meaningful help when it was private than when it was public (Journal of Consumer Research 40(6)). A public like or share can satisfy the wish to be seen as supportive. So the ask that matters, the hearing, the comment letter, the signup, should come close behind the token, and should be framed as the thing the person already committed to, not as a new favor.

Posting frequency and timing

Benchmark data on frequency is plentiful and points one way. Socialinsider's 2025 averages: brands posted about 24 times a month on Facebook, 20 on Instagram, 15 on TikTok and 70 on X, and average engagement fell year over year on every platform (Socialinsider). Rival IQ recorded median posting frequency slowing on Facebook, dipping on Instagram for the first time in three years, and falling 33 percent on X, while engagement fell by a third or more on most platforms (Rival IQ). Nothing in either dataset shows a reward for volume; both show high-posting brands with low per-post engagement. For an advocacy page the right reading is that cadence is an operations decision (what you can sustain with sources checked, see lesson 10), and the benchmark figure to track is engagement per post, not posts per week.

Headlines and the curiosity gap

The largest randomized dataset on headlines is Upworthy's. The site tested headline variants against each other for years, and Robertson and colleagues analyzed the archive: 22,743 randomized controlled trials, about 105,000 headline variants, about 5.7 million clicks and over 370 million impressions in the confirmatory sample. Each additional negative word in an average 15-word headline raised click-through by about 2.3 percent; each positive word lowered it by about 1.0 percent (Nature Human Behaviour, 2023). The effects are small per word and consistent, and they compound. The Journalism site's lesson on reading the numbers warns about what clicks do and do not measure (reading the numbers); a click is not a changed mind, and a headline that earns the click with a word the story does not justify produces the retention cliff described above.

The curiosity gap draws on Loewenstein's 1994 account of curiosity as a response to a perceived gap in one's knowledge (Psychological Bulletin 116(1)). The legitimate form names the question the reader did not know they had and promises the answer: "The water district's fixed charge is going up 40 percent. Page 12 of the filing explains why." The clickbait form withholds the subject entirely: "You won't believe what's on page 12." The first survives the click; the second earns it and loses the reader by second four. Robertson's numbers say negativity sells; Green and Brock's say absorbed readers update beliefs regardless of truth. Put together, they describe exactly the temptation an advocacy page should refuse, for reasons that lesson 01 lays out: the outrage economy is real, and a page that declines to run on it has to win on receipts instead.

Testing on a small page

Almost every content claim above ends with "test it on your own page." Here is why most such tests fail, and how to run one that does not.

An A/B test compares two versions on randomly assigned audiences and asks whether the difference in some rate (clicks per impression, engagement per reach) is larger than chance would produce. Four numbers decide how many observations you need: the baseline rate, the smallest difference you care to detect (the minimum detectable effect), the significance level (usually 5 percent, the chance of a false positive) and the power (usually 80 percent, the chance of catching a real effect of that size). Evan Miller's calculator takes exactly those inputs; its own example is a 10.2 percent baseline and a lift to 13.2 percent needing 2,545 subjects per variation (Evan Miller). The arithmetic behind it is the standard two-proportion formula, and the results for the rates an advocacy page actually sees are sobering.

What you want to detect	Baseline	Target	Observations needed per variant (5 percent significance, 80 percent power)	Realistic?
Caption style lifts engagement per follower	0.50 percent	0.60 percent	About 86,000	Not on a page reaching 2,000 per post
Caption style doubles engagement	0.50 percent	1.00 percent	About 4,700	Yes, over a few posts, if the effect is that big
Image versus carousel on link clicks	1.0 percent	1.5 percent	About 7,750	Over a month on a mid-size page
Share rate on document posts	3.0 percent	6.0 percent	About 750	Yes, quickly
Advocacy email click rate (M+R average 2.3 percent)	2.3 percent	3.0 percent	About 8,300	On a list of 4,000, two sends per arm
Email open rate (Mailchimp overall 35.6 percent)	35 percent	38 percent	About 4,000	Yes, but opens are unreliable after Apple Mail Privacy Protection (lesson 07)

The pattern: small differences on low-rate outcomes need tens of thousands of observations, which a local page does not have. Large differences on higher-rate outcomes (shares on receipts, clicks on an email) need hundreds to a few thousand, which it does. So test big changes, on the outcome closest to the action you want, and accept that fine-tuning (emoji or not, 9 a.m. or 10) is below your page's resolution.

Questions and answers

Our reels get far more reach than our document posts, but the document posts get the comments and shares. Which should we make more of?

Both, for different jobs, and measure the one that leads to action. Reach from Reels comes largely from recommendation to people who do not follow you; Instagram's ranking explainer says Reels and Explore weigh popularity signals heavily and predict reshares and completion (Instagram). Document posts earn engagement from people who already care, which the Rival IQ nonprofit data on Facebook photos (the top-engaging format for nonprofits) supports. Lesson 08 builds the funnel; the usual finding is that reels bring new followers and documents move existing followers to act.

Is a 0.5 percent engagement rate bad?

On Instagram in 2025 it was the brand average (0.48 percent across 35 million posts) and on Facebook it would be three times the average (0.15 percent) (Socialinsider). Rival IQ found nonprofits above median everywhere. Compare yourself with the platform average, the nonprofit figure and, most of all, your own trend. A rate that falls as followers grow is normal arithmetic, since the denominator grew; a rate that falls with flat followers is a content problem.

Should we post more to fight falling reach?

The benchmark data gives no reason to think so. Brands posted 20 to 24 times a month on Instagram and Facebook in 2025 while engagement fell 16 to 36 percent year over year across the two big benchmark sets (Socialinsider; Rival IQ). More posts with the same quality means the same engagement spread thinner, and on a page network it invites the duplicate-content penalties in lesson 04. Post what you can source and sustain.

Do hashtags help?

Instagram's 2023 explainer does not list hashtags among its ranking signals; TikTok's 2020 explainer lists captions, sounds and hashtags as "video information" signals, below user interactions in weight (TikTok). Rival IQ reports which hashtags nonprofits used most (awareness-day and cause tags), which tells you about habit, not effect. We found no randomized evidence this session that hashtag count changes reach. Use the few that describe the post. The beginner course covers the practice (anatomy of a post).

Our best-performing post ever was angry. Doesn't that prove outrage works?

It proves one post performed, and the arousal research says anger is one of the high-arousal emotions that spread (Berger and Milkman). The same research found awe spreads too, and Brady's data says moral-emotional language spreads mostly inside your own group (Brady et al.). An angry post reaches your own side. A surprising document reaches people who had not heard. Lesson 01 is about choosing which high-arousal response you are willing to run on.

How long should a video be?

Long enough to keep the promise and no longer. Facebook describes reels as "often 90 seconds or less" and says longer videos still have a place (Meta for Creators). TikTok weighs finishing a longer video as a strong signal. YouTube's retention report will show you where your audience leaves. The reels course has the platform limits as of its last check; cut to the length the story earns.

Can we use AI-generated images for posts about real events?

Not for anything presented as a depiction of a real event or person; lesson 09 covers Idaho's synthetic media election law and platform labeling rules. For illustrative images, label them. The persuasion evidence in this lesson all concerns real accounts and real documents; a receipt that turns out to be generated is the end of a page's credibility.

Our test showed a 12 percent lift but the calculator says we needed four times the sample. Can we use it?

Treat it as a lead, not a result. An underpowered test that shows a lift is as likely to be noise as signal, and the lifts that survive tend to shrink on replication. Run it again at the computed size. If it holds, you have something; if it does not, you saved yourself a year of production time on a format that did nothing.

Hard questions

Isn't all of this just manipulation with footnotes?

The techniques here are the same ones every communicator uses: tell a true story, show the document, open with the point, ask plainly. What the research adds is a size for each effect and a warning about the biggest one. Green and Brock found that absorbed readers update beliefs whether a story is labeled fact or fiction, which is a reason to be strict about truth, not permission to be loose with it. The line this course draws, in lesson 09, is that every persuasive technique is fair when the underlying claim is true and sourced, and none is when it is not. Manipulation is deceiving people about facts. Persuasion is arranging true facts so people attend to them.

The data says negativity and anger spread. Why shouldn't an advocacy page use what works?

Because the same data shows where it goes. Brady's 563,312 tweets diffused mostly within ideological groups; Robertson's 2.3 percent per negative word buys clicks, not persuasion; and Rathje and colleagues' PNAS 2021 study, covered in lesson 01, found out-group animosity the strongest predictor of sharing. A page that runs on outrage reaches the already convinced and hardens the rest. For an effort whose theory of change needs a council vote or a majority, that is the wrong audience. Where a fact is genuinely alarming, say so; the alarm is earned. Manufacturing it is both unethical and, by the group-boundary evidence, ineffective for the people you most need.

Benchmarks from brands with marketing departments tell a volunteer page nothing. Why cite them?

Because they are the only large-sample descriptive data that exists, and because the alternative is folklore. The lesson treats them as a prior to test, not a conclusion, and spends its last section on how to test them at your own scale. Where the benchmark says carousels out-engage images by half, a small page can check that in six posts. Where a benchmark says 9 a.m. Tuesday, no page can check it, and the lesson says so.

If most A/B tests on small pages are underpowered, isn't the honest advice to stop testing?

The honest advice is to test only what you can detect. Large effects on high-rate outcomes (doubling share rate on document posts; a 50 percent lift in link clicks) are within reach of a page with a few thousand followers over a few weeks, and those are the effects worth having. Fine-tuning is not, and pretending otherwise produces the churn of "optimizations" that never replicate. One well-powered test a quarter beats a weekly stream of noise.

Doesn't the story research cut against the "post the document" advice? People are moved by narrative, not budget pages.

The effect sizes do not support that opposition. Narratives move beliefs at about r = 0.17; the useful and surprising content Berger and Milkman found more viral is often a document. The combination, a person's account anchored to the page that proves it, uses both. A page that posts only stories asks to be trusted; one that posts only documents asks to be read. Advocacy needs both, and the receipts are the part opponents cannot take away.