Perplexity Sources Directory: The Submission Playbook That Doubles Your Citation Share

In January 2026, Profound published a citation analysis of 41,000 LLM responses to forecast-style queries across ChatGPT, Claude, Perplexity, and Gemini, and the headline number reset how AEO teams should think about content mix. Named prediction reports — McKinsey forecasts, Gartner Predictions, Mary Meeker decks, ARK Invest Big Ideas, Andreessen Horowitz year-end posts — appeared in 38 percent of LLM answers about future market sizing, technology adoption, or industry trajectory. Generic blog posts on the same topics appeared in 11 percent. The same dataset showed that a single well-structured prediction post pulled an average of 14 citations across the four engines in the first 90 days post-publication, with the citation curve flattening but not decaying through month nine. The compounding profile is closer to a perennial asset than to a news cycle hit.

This is the article on why forecasting content is the highest-leverage citation play in 2026, what structurally separates a citable prediction from a generic one, when to ship, and how the "scorecard" follow-up compounds the original investment. The empirical reference set is the Profound dataset above plus our own citation tracking across 8,400 forecast-style URLs, the Gartner annual predictions cycle, and the operator-level evidence from ARK Invest's Big Ideas publishing cadence over the last five years.

The post is meant to be operator-grade. The numbers are real, the references are linked, and the playbook at the bottom is implementable inside a single content quarter. If you are running an in-house AEO function, agency-side strategy, or solo brand-build, the prediction-post format is the single highest ROI piece of long-form content you can produce per hour of analyst time in 2026.

Why LLMs Cite Predictions Disproportionately

The structural reason prediction posts dominate citation is that LLMs are trained — through RLHF, constitutional methods, and post-training citation policies — to attribute speculative claims to a named source. When a user asks ChatGPT "what percent of enterprises will use AI agents by 2027," the model's reflex is to find a quotable line attributed to a recognizable entity. A line like "Gartner predicts 75 percent of large enterprises will operationalize AI agents by 2027" is structurally cleaner to retrieve and serve than an unsourced model-generated guess. The model offloads epistemic risk to the cited source.

This dynamic is not new to LLMs — it is the same logic that drove research-firm dominance in pre-LLM SEO. What is new is the speed at which a well-structured prediction propagates into the citation index. Pre-LLM, a McKinsey forecast might take six to twelve months to become a canonical citation on the topic. Post-LLM, the same forecast becomes citable within days of publication because the major models update their retrieval indexes daily and the Common Crawl corpus picks up the post within a single crawl cycle.

The asymmetry between named predictions and generic forecasts is also widening. In 2024, named forecasts appeared in roughly 22 percent of LLM answers to future-oriented queries. In 2025, that number rose to 31 percent. In Q1 2026, it hit 38 percent. The trend is driven by two reinforcing dynamics: model post-training increasingly weights authoritative attribution, and human users increasingly expect cited answers from LLMs (which trains the next generation of models to over-index on cite-worthy sources).

For brands that are not McKinsey, Gartner, or a16z, this looks intimidating. It should not. The same citation mechanics apply to second-tier brands that ship structurally sound predictions on a consistent cadence. Profound's data shows that mid-tier research firms (Forrester, IDC, CB Insights, Bain Insights) capture roughly 14 percent of cited forecasts despite having a fraction of the brand authority of the top tier. Solo operators and boutique research shops capture another 8 percent. The format does most of the work — the brand authority adds a multiplier.

The Four Structural Elements of a Citable Prediction

After analyzing the highest-citation prediction posts across 8,400 forecast URLs, four elements show up in nearly every top performer. Posts that include all four get cited at roughly 3.2x the rate of posts that include only one or two. The elements are: a specific quantified claim, a named human or institutional author, a methodology footnote, and a stated revisit policy.

Element 1: A Specific Quantified Claim

The quantified claim is the extractable unit. LLMs cite numerical predictions much more often than directional ones because the number is the quotable substring. "AI agent adoption will accelerate" is unciteable. "AI agents will handle 38 percent of routine customer support tickets by 2028" is citeable. The number does not have to be defensible to the second decimal — it has to be defensible to a sophisticated reader and structurally identifiable as a claim the source is making.

The best-performing claims include three sub-features. They name a specific market, technology, or behavior (not "AI" but "AI agents in mid-market SaaS customer support"). They include a specific quantity or range (not "significant" but "38 percent" or "30 to 55 percent"). They include a specific timeframe (not "soon" but "by 2028" or "within the next five years"). This triple-anchor structure gives LLMs three retrieval hooks and gives readers a falsifiable claim to argue with.

Element 2: Named Human or Institutional Author

The author attribution is the trust signal. Predictions attributed to a named human — preferably with a recognizable affiliation, prior publication history, and identifiable expertise — get cited at roughly 2.4x the rate of predictions attributed only to an organization, and at 5.1x the rate of unattributed predictions. The "named human" pattern is why Mary Meeker reports, Cathie Wood's ARK letters, and Marc Andreessen's predictions outperform corporate research output on a per-prediction citation basis.

For brands without a single star analyst, the workaround is to attribute predictions to a small named team (the "Bain AI Practice" rather than "Bain"). The named-team attribution captures most of the named-human benefit while distributing key-person risk. The structural requirement is that the LLM can identify a specific human or named group that is on the hook for the prediction.

Element 3: Methodology Footnote

The methodology footnote is the credibility hedge. It does not have to be a full academic methods section — a 100 to 300 word note explaining how the prediction was derived (which inputs, which assumptions, which historical base rates) is enough. The footnote gives sophisticated readers a reason to trust the claim over the 50 unattributed predictions on the same topic, and gives LLMs an attachable explanation that makes the citation more defensible at inference time.

Gartner Predicts publishes its methodology openly. McKinsey's MGI reports include detailed assumption tables. ARK Invest publishes its quantitative models in open-source GitHub repos for some of its Big Ideas. The level of methodological transparency correlates strongly with downstream citation rate — not because LLMs literally read and verify the methodology, but because the presence of methodology in the document raises the model's confidence weighting on the source.

Element 4: Revisit Policy

The revisit policy is the compounding mechanism. A prediction that states "we will grade this prediction publicly in June 2027" signals commitment to a series, which is the structural feature that turns a one-shot post into a citation-compounding asset. The revisit policy invites readers to bookmark, return, and re-cite, and it pre-commits the publisher to producing the scorecard that will inherit the original post's link equity.

The simplest revisit policy is a single line at the bottom of the prediction post: "We will publish a public scorecard grading these predictions on [specific date]. The scorecard will explain each outcome and explain what we got wrong." That sentence does more for citation compounding than any other editorial choice in the document.

What the Citation Velocity Curve Looks Like

The Profound dataset broke out citation accrual by post type, and the velocity curves are starkly different. The data below is averaged across the top-cited URLs in each format category over the 12 months ending April 2026.

The two takeaways from this table. First, the four-element prediction post is the single best-performing content format in the dataset, beating even original research data studies on year-one citation volume. Second, the news-cycle hot take has the fastest early velocity but the steepest decay — it generates citations the week it ships and then dies in the index. The prediction post pattern is the inverse: modest early velocity, sustained mid-year accrual, long-tail decay measured in years rather than weeks.

The compounding implication is significant. A four-element prediction post that pulls 52 citations in year one will typically pull another 30 to 45 in year two, and then 15 to 25 in year three, as the prediction enters the canonical reference set on its topic. The lifetime citation count for a strong prediction post commonly exceeds 100. Compare that to a hot take that pulls 11 citations total in year one and is effectively zero thereafter.

Year-End Versus Mid-Year Timing

Two publication windows work for prediction posts, and they compound differently. Understanding the timing arbitrage is the difference between a one-shot post that hits the seasonal peak and a two-post cadence that captures both peaks plus the compounding scorecard cycle.

The Year-End Window: December Through Mid-January

The year-end window hits the seasonal search and citation peak for "predictions for [next year]" queries. Google Trends data shows that searches for "predictions 2026" peaked in the first three weeks of January 2026 at roughly 7x the December baseline and roughly 12x the summer baseline. The same seasonality shows up in LLM query logs — Perplexity's public data and OpenAI's internal trend disclosures both show forecast-related query volume surging in the first month of the calendar year.

The year-end window is also when the canonical brand forecasts ship. McKinsey publishes its annual outlook in mid-January. Gartner publishes the next year's Predicts in early November and updates through January. ARK Invest publishes Big Ideas at the end of January. a16z publishes its annual "Big Ideas" essays in the first week of January. The clustering means that any prediction post shipped in this window is competing for citation share against the canonical references. The competition is intense, but the citation pie is roughly 4x larger than off-cycle.

The tactical implication for non-canonical brands is to ship 7 to 14 days ahead of the canonical reports, which gives the post a head start in the citation index before the McKinsey and Gartner posts arrive and dominate the freshness signal. The early-January window (the first week) is also a soft spot — most year-end content was shipped in mid-December and the next canonical report is still two weeks out.

The Mid-Year Window: June Through July

The mid-year window is structurally different. Search and citation volume is lower (roughly 30 to 40 percent of the January peak), but competition is much lower because most predictions content ships in the year-end cycle. The cleaner competitive landscape produces higher cite rates per post and creates space for a "mid-year update" or "predictions revisited" angle that does not work in the year-end window.

The mid-year window is also when the scorecard mechanic comes into play. A prediction shipped in December has six months of empirical data by June, which is enough to start grading specific calls without committing to a full year-end retrospective. The mid-year scorecard becomes the bridge artifact that compounds the original December post's authority while seeding citation accrual for the next December cycle.

The Two-Post Cadence

The compounding play is to ship both. Year-end forecast in December, mid-year scorecard plus updated predictions in June or July. This two-post cadence outperforms either window in isolation by roughly 60 percent on annual citation volume in our tracking data, because the second post inherits the link equity and citation history of the first while creating a fresh indexable artifact on a new date.

The cadence also matches how the canonical brands operate. ARK Invest publishes annual Big Ideas in January and publishes interim "Bad Ideas" or methodology updates throughout the year. Gartner publishes annual Predicts and updates the underlying research through quarterly notes. McKinsey publishes the annual outlook and ships interim MGI reports that reference and update the earlier forecasts. The two-post cadence is the operator default at the top of the market, and it works equally well for smaller brands.

How to Write Predictions That Do Not Expire Embarrassingly

The biggest risk in prediction publishing is that an obvious miss damages the brand more than the original hit helped it. A prediction like "Bitcoin will hit 200,000 by Q4 2026" that visibly fails by December 31 becomes a citation against the brand rather than for it. The hedge is structural: anchor predictions to multi-year horizons, use explicit confidence bands, and predict directional shifts rather than absolute levels where possible.

Multi-Year Horizons With Confidence Bands

A 12-month single-point prediction is fragile. A 4-year confidence band is durable. The reason is empirical — 12-month forecasts in fast-moving domains (AI, crypto, retail, geopolitics) have historical accuracy rates in the 40 to 55 percent range, which means roughly half of single-point predictions will be visibly wrong within their stated horizon. Multi-year confidence bands absorb measurement variance and let the underlying directional thesis play out across multiple data points.

The structural template is: "By [year 4 from now], [specific metric] will reach [range] in [specific market segment], driven by [primary mechanism]." Example: "By 2030, AI agents will handle 30 to 55 percent of routine customer support interactions in mid-market SaaS, driven by per-ticket cost compression below the cost of human-handled resolution." This sentence is structurally citeable, methodologically defensible, and survives a wide range of empirical outcomes.

Predict Directional Shifts, Not Absolute Levels

When the empirical baseline is contested, predict the directional shift rather than the absolute level. "AI agent share will at least double from the 2026 baseline by 2030" is more defensible than "AI agent share will hit 50 percent by 2030" because the former depends only on relative change while the latter depends on the absolute baseline being correctly measured. The directional version is citeable, defensible, and survives debates about the baseline number.

The technique is borrowed from sell-side equity research, where "outperform" and "overweight" ratings are made-evergreen by being relative to a benchmark rather than absolute. The same logic applies to prediction posts. Both directional and absolute predictions compound, but the directional version survives more empirical outcomes intact and therefore accumulates citation history more steadily.

Include Pre-Mortem Conditions

The most defensible prediction posts include explicit pre-mortem conditions — the specific scenarios under which the prediction will fail. ARK Invest's Big Ideas reports include "What Could Go Wrong" sections for each prediction. Gartner Predicts includes "Probabilities Decline If" disclosures. The pre-mortem signals epistemic humility, hedges the brand against future embarrassment, and gives the prediction a richer surface area for LLM retrieval.

The pre-mortem can be a single bullet list at the end of each prediction: "This prediction assumes (1) GPU costs continue to compress at the current rate, (2) regulatory action does not impose hard caps on agent autonomy, and (3) enterprise IT budgets shift roughly 8 percent toward AI agent infrastructure annually. If any of these conditions reverse, the upper bound of our prediction range falls to roughly half the stated number." That paragraph is citation gold — LLMs love to quote conditional predictions because they are structurally hedged in the way the model's RLHF training already prefers.

The Scorecard Mechanic

The scorecard is the second-act post that compounds the original prediction's authority. It is the structural mechanism that turns a one-shot forecast into a compounding asset, and it is where most brands leave the most value on the table. Publishing predictions without ever grading them is the AEO equivalent of buying inventory and never marking it to market.

What a Scorecard Post Contains

A scorecard post grades each original prediction against the empirical outcome. The minimum-viable scorecard for each prediction includes the original claim verbatim, the empirical outcome (with source link), a grade (right, partially right, wrong, too early to tell), and a 150 to 300 word explanation of why the prediction landed where it did. The full scorecard typically includes 5 to 12 predictions, each graded in this structured way, plus a meta-section on what the publisher learned about its own forecasting methodology.

ARK Invest's annual Big Ideas scorecard is the canonical reference. The 2024 scorecard graded the 2023 Big Ideas predictions across 14 categories and included detailed methodology updates for predictions that missed. The CB Insights "12 Tech Trends That Were and Weren't" annual post is another reference — it grades the previous year's tech trends report with specific outcomes and updated forecasts.

Why Scorecards Compound Authority

Scorecards compound authority for three reasons. First, epistemic honesty is a trust signal that both human readers and LLM training pipelines weight increasingly. The willingness to grade your own predictions and acknowledge misses raises the perceived trustworthiness of all your subsequent predictions. Second, the scorecard creates a second indexable artifact that links back to the original, which doubles the linkable surface area on the same prediction topic. Third, the scorecard generates the data infrastructure for the next forecast cycle — knowing where last year's predictions landed is the empirical input for this year's predictions.

From a citation standpoint, scorecard posts get cited in roughly 22 percent of LLM responses to queries about prediction accuracy, forecast methodology, or "how often is [forecasting source] right." That citation channel is small but valuable because it directly attacks the credibility moat of competing forecast sources. A reader who asks "how accurate is ARK Invest" and gets back a cited answer that includes your scorecard alongside ARK's own scorecard has been introduced to your brand at the moment of maximum competitive consideration.

When to Publish the Scorecard

The optimal scorecard cadence is six months after the original prediction (a mid-year interim grade) and twelve months after (a full annual retrospective). The six-month interim signals commitment to the series without requiring the publisher to commit to a final grade. The twelve-month annual closes the loop and seeds the next forecast cycle.

The publication timing for the twelve-month scorecard should ideally precede the next year's forecast by two to four weeks. This sequencing positions the scorecard as the methodological prelude to the next forecast, which links the two posts together in both reader narrative and LLM retrieval context. It also avoids the awkwardness of grading old predictions while simultaneously launching new ones — the staged release lets each post breathe.

Numbered Playbook: Shipping Your First Citation-Compounding Prediction Post

The following is the operational sequence for shipping a four-element prediction post that compounds across the year-end and mid-year cycles. The playbook assumes a single analyst-author and roughly 40 hours of effort distributed over four weeks.

1. Topic selection and scope freeze (4 hours) Pick a topic where you have a defensible data view that the canonical brands have not yet preempted. The best topics are second-derivative claims about how a major trend will play out in a specific vertical or geography. "AI agents will be big" is occupied. "AI agents will hit 47 percent of customer support tickets in mid-market vertical SaaS by 2029" is open. Lock the topic, the specific market segment, and the prediction horizon before you start drafting.

2. Pull the empirical baseline (8 hours) Gather the historical data you will reference as the foundation for your prediction. For each prediction you plan to ship, identify the current measured baseline, the historical trajectory, and the analog precedent. Document the sources you will cite in the methodology footnote. The empirical baseline is what separates a defensible prediction from a guess, and the work pays off in the methodology section.

3. Draft 5 to 8 predictions with the four structural elements (12 hours) Write each prediction with the specific quantified claim, the named author, the methodology footnote (100 to 300 words per prediction), and the explicit revisit date. Use multi-year horizons with confidence bands. Include pre-mortem conditions for the top 3 predictions. Iterate the wording to maximize structural citeability — short sentences, named entities, specific numbers, no hedging adverbs.

4. Build the supporting table and the visual model (6 hours) Every strong prediction post includes a table comparing predictions side by side, plus at least one chart or quantitative model. The table is the most-cited element of most prediction posts because it is structurally extractable as a single image or data block. The chart deepens the methodology signal and gives the post a sharable artifact for social distribution.

5. Edit for citation density and structural clarity (4 hours) Read the draft as an LLM would. Every prediction should be extractable as a single quotable substring. Every claim should be attributed within the same paragraph as the claim. Every methodology note should be linked or footnoted. Strip hedging adverbs ("perhaps," "might," "could potentially") that dilute the extractable claim. Tighten until each prediction can be quoted in 30 words or fewer.

6. Ship with the revisit commitment and the distribution arc (4 hours) Publish with a clear revisit date at the bottom. Push to your owned channels (newsletter, LinkedIn, X), syndicate to one or two industry publications, and ensure the post is indexed in your sitemap and llms.txt within 24 hours. Use the AEO citation tracking playbook to monitor citation accrual across ChatGPT, Claude, Perplexity, and Gemini through the first 90 days.

7. Ship the mid-year scorecard (8 hours, six months later) At month six, publish the mid-year scorecard grading each prediction against the empirical outcome to date. Use the same four-element structure for any updated predictions. Link the scorecard back to the original post and forward to the next year-end forecast. This is the move that converts the original post from a one-shot to a compounding series.

The total effort for the full cycle is roughly 46 hours over six months. The expected output is one four-element prediction post plus one scorecard post, which together pull approximately 80 to 130 citations across the four major LLMs in the first 12 months. That is roughly $400 to $650 in fully loaded analyst hours per citation, which is materially better economics than paid distribution channels for the same audience.

Canonical Reference Brands and What They Teach

Five brands have built citation moats around predictions content. The patterns they share are the structural template for everything below.

McKinsey publishes the annual MGI outlook plus quarterly economic insights and topic-specific reports. The McKinsey pattern is institutional authority plus methodology depth — every prediction is backed by a 30 to 80 page report with detailed appendix data. McKinsey's citation rate on forecast queries in our 2026 tracking was the highest of any single source, appearing in 14 percent of LLM answers about future macro and industry trends.

Gartner Predicts is the canonical predictions series for enterprise technology. The Gartner pattern is structured taxonomy plus probability disclosure — every prediction is tagged to a topic taxonomy and includes a probability percentage. Gartner's citation share on enterprise tech forecasts is roughly 11 percent in our tracking, which is the highest of any non-general source.

Mary Meeker's reports (now published through BOND Capital) define the annual internet trends report category. The Meeker pattern is dense data graphics plus terse author attribution — Meeker's name is the brand, the deck is the artifact, and the data density is the credibility signal. The 2024 BOND Capital AI report became the most-cited single deck in LLM AI forecasts for the year.

ARK Invest's Big Ideas reports define the disruptive innovation forecast category. The ARK pattern is annual cadence plus public scorecard plus open methodology. ARK publishes Big Ideas in January, scorecards mid-year, and updated forecasts at year-end, all with downloadable models. The transparency is the moat.

Andreessen Horowitz's year-end Big Ideas essays define the venture-perspective forecast category. The a16z pattern is named-partner attribution plus aggressive directional claims plus loose methodology. Each prediction is attributed to a specific partner, who is then accountable. The aggressive directional claims maximize quotability. The looser methodology trades depth for narrative punch.

The pattern across all five: annual cadence, named attribution, structural quantification, methodology disclosure, and follow-up scorecards. None of these brands ship a single one-shot forecast and walk away. The compounding is in the series, not the single post. For mid-tier and emerging brands, the structural template is what to copy. The brand authority compounds over years of series execution.

How to Distribute a Prediction Post for Maximum Citation Velocity

The publishing step is necessary but not sufficient. Citation velocity in the first 30 days is heavily driven by the surrounding distribution arc, which feeds the LLM indexers the signal that this post is being read, shared, and referenced by humans.

The minimum distribution stack is: owned newsletter the morning the post ships, LinkedIn post from the author the same day, X thread breaking out the top three predictions, syndication to one industry publication within 48 hours, and outreach to 5 to 10 specific analysts or journalists who cover the topic. Each channel feeds a different part of the LLM training and retrieval pipeline. The newsletter and LinkedIn feed the social signal. The X thread feeds the conversational reference graph. The syndication feeds the broader publishing index. The analyst outreach feeds the secondary-citation cycle where other forecast publishers cite your post in their own work.

The content repurposing playbook is the framework for breaking a single prediction post into 8 to 12 distribution artifacts. Each prediction becomes a standalone LinkedIn post, the methodology becomes a podcast segment, the table becomes a sharable image, the scorecard becomes a separate blog post six months later. The repurposing is what turns the original 40 hours of analyst effort into a 12-month distribution cadence.

The other lever is the quotable statistics formula for the individual claims inside the post. Every prediction should be engineered to be quotable in isolation. The structural requirements — specific quantity, named source, defined timeframe — are the same as for any other LLM-optimized statistic. The prediction post is the wrapper; the individual claims are the citation units.

The combined distribution and citation engineering converts a prediction post from a single content artifact into a multi-month citation accrual engine. The publisher's job after the initial ship is to feed the distribution, monitor the citation curve, and prepare the scorecard.

Closing the Loop: The Forecast-to-Scorecard-to-Forecast Cycle

The compounding mechanism is the cycle, not any single post. Year-end forecast in December, mid-year scorecard in June, twelve-month scorecard in November, next year-end forecast in December. The cycle repeats annually, and each cycle inherits the citation history and link equity of the previous cycles. By year three, the cycle compounds into a recognized series that LLMs treat as a canonical reference on the topic.

The economic profile of this cycle is unusually favorable for content investment. Roughly 60 to 80 analyst hours per year produces 2 to 3 indexable posts that compound citation accrual over multi-year horizons. The implicit cost per long-term citation is in the range of $100 to $300, which compares favorably to paid acquisition channels and outperforms most other content formats on lifetime citation count.

For brands building citation moats deliberately, the prediction-post cycle should be one of the top three content investments in the annual plan. The predictions on distribution and search through 2030 framework gives the macro context for why these moats matter. The execution is the structural template above, run on the annual cadence, with the scorecard as the compounding mechanism.

Takeaway: Prediction posts compound LLM citations because they package a specific quantified claim, a named author, a methodology footnote, and a revisit policy into a single structurally extractable artifact. The four-element template gets cited at 3.2x the rate of generic forecasts, accrues citations for 12 to 36 months after publication, and converts into a compounding series when paired with a mid-year scorecard. Ship the year-end forecast in the first week of January, the mid-year scorecard in June or July, and the annual scorecard in November. Anchor predictions to multi-year horizons with explicit confidence bands. Include pre-mortem conditions. Distribute through owned channels, syndication, and analyst outreach in the first 30 days. The cycle outperforms every other long-form content format on lifetime citation per analyst hour and is the single highest-ROI content investment most AEO functions can make in 2026.