FAQ & Methodology

1. What is the Intelligence Price Index?

The Intelligence Price Index (IPI) tracks the revealed price of AI-exposed knowledge work over time. It is a matched-model price index built from the posted prices of freelance "gigs" on Fiverr — tasks like logo design, copywriting, coding, voice-over, and video editing — that generative AI can increasingly perform.

The motivating question is simple: as AI gets better at cognitive tasks, what happens to the market price of doing those tasks? Rather than surveying experts or estimating exposure scores, the IPI reads the answer directly off prices that real sellers set and revise in a competitive marketplace. It is built like the Consumer Price Index (CPI), but the "basket" is cognitive labor instead of groceries and rent.

2. Who is this for?

The IPI is built for people who work on or hire through freelance marketplaces — primarily Fiverr and Upwork — and for researchers tracking how AI is repricing knowledge work.

• If you sell logo design, copywriting, coding, voice-over, translation, or video editing, the index is a read on where the going rate for your kind of task has been heading.
• If you buy that work, it is a read on what you should expect to pay.

The numbers are drawn from one marketplace (Fiverr, see below), but the tasks priced are the same standardized gigs that dominate both platforms.

3. What does the headline number mean?

The big number at the top of the index is the change in the composite index over the full charted period — from the 2020 Q1 base to the latest quarter (labelled Δ'20–'26 on the page). Each category and the composite are set to 100 at the base quarter, so a reading of, say, −5% means the price of the selected basket of AI-exposed work is about 5% lower than it was at the start of 2020.

• A negative number = deflation — the work is getting cheaper (substitution by AI tends to push this way).
• A positive number = inflation — the work is getting more expensive (general inflation or AI complementing rather than replacing the worker).

The chart shows the full quarterly path of each category and the composite; the table ranks categories by their full-period change, steepest decline first. You can also read shorter movements straight off the chart between any two quarters.

4. What time period does the chart cover, and why does it start in 2020?

The chart plots the full quarterly series from 2020 Q1 to 2026 Q1 — 25 quarters, with the index fixed at 100 in 2020 Q1. The underlying archive reaches back to ~2011, but the chart opens in 2020 for two reasons:

• Coverage. Before 2020 the archive is thin and uneven — too few matched gigs per quarter to draw a stable line.
• A clean baseline. 2020 gives a pre-generative-AI baseline (ChatGPT launched November 2022) while keeping the window dense enough to trust.

Earlier history may be added as coverage improves. We show quarters rather than months because quarterly buckets contain more matched gigs each period, so the series is far less jumpy than a monthly one.

5. Is the effect of ChatGPT (or COVID) visible in the chart?

The window is deliberately drawn so you can look. ChatGPT launched in November 2022 (2022 Q4), so the quarters on either side of that line are the natural place to look for an AI-substitution signal in exposed categories like writing and design.

The chart also opens during the COVID shock (2020–2021), which moved freelance demand for its own reasons — so be careful reading the earliest quarters as an AI story. The IPI shows the price movement; it does not, by itself, prove what caused any given move. Separating the AI effect from COVID, macro inflation, and platform growth is exactly what the accompanying paper works on — see the limitations below.

6. What exactly is being priced?

The unit of observation is a gig: a single, well-defined task a seller offers at a posted price (e.g. "design a minimalist logo," "translate 500 words EN→ES"). We use the Basic price tier. Three properties make these prices well-suited to an index:

• Revealed, not surveyed. Sellers set these prices in response to competition — they are market outcomes, not estimates.
• Standardized tasks. A gig is a comparable bundle of work, so its price can be tracked over time like a CPI item.
• Matched over time. We only compare a gig's price to its own earlier price, so we measure genuine price movement, not changes in which sellers happen to be in the sample.

7. How were the categories chosen, and what counts as "AI-exposed"?

The index currently tracks seven categories — design, writing, marketing, coding, video, audio, and translation — chosen on two tests:

• A standardized, posted-price gig that recurs across many sellers, so it can be matched to its own price over time.
• Plausibly AI-exposed — something generative AI can increasingly do or assist.

Categories are assigned by classifying each gig from its archived page. Work that is mostly manual, in-person, or not cleanly priced as a fixed gig — data entry, virtual assistance, admin support, general "consulting" — is not tracked, because it either resists standardized pricing or is not AI-exposed in the same way. The category set is meant to grow; it is not a claim that these seven are the only AI-exposed work.

8. Where does the data come from?

Every price on this site is a real, historical Fiverr list price recovered from the Internet Archive's Wayback Machine — the public web archive that has been snapshotting Fiverr gig pages since around 2011. Nobody is surveyed and nothing is estimated: we simply read the prices sellers actually posted, as preserved in old archived copies of their pages, and line them up over time.

From 60 million archived URLs to a clean price panel

The data is built through a multi-stage pipeline that narrows a vast pile of raw archive entries down to a set of gigs whose prices can be tracked reliably:

How a price is read off each page

Fiverr's page layout changed a lot over the years, so extraction uses a cascade of four methods, trying the most reliable first and falling back as needed:

What this site shows specifically

The full study spans 2011–2026. The figures on this page use the full-history quarterly build — the pipeline aggregated into quarters and charted from 2020 Q1 to 2026 Q1 (25 quarters). The number of matched gigs behind each category is shown in the Gigs column on the index page. Because this is a pilot-scale sample, please read the limitations before drawing strong conclusions from any single category.

9. Why not just survey freelancers or use official wage data?

Surveys ask people what they think prices are doing; the IPI reads what sellers actually posted. Revealed prices cannot be colored by recall, sentiment, or who happened to answer.

Official wage statistics (e.g. BLS series) are real too, but they are aggregated, lagged, and not broken out by the specific AI-exposed gigs we care about — you cannot see the price of "a minimalist logo" or "500 words EN→ES" in them. Fiverr's posted, packaged prices give exactly that: a task-level list price you can match to its own past, quarter after quarter. The trade-off is coverage (one marketplace, posted not transacted), which we are upfront about in the limitations.

10. Why Fiverr — and what about Upwork?

Two practical reasons.

• Posted, packaged prices. Fiverr gigs carry explicit, fixed price tiers (Basic / Standard / Premium) that sellers set up front. That makes each gig a clean, comparable price point you can track quarter to quarter — exactly what a price index needs. Upwork work is mostly negotiated hourly or per-project, so there is no single posted "list price" to follow over time.
• A deep archive. The Wayback Machine has been snapshotting Fiverr gig pages since ~2011, giving a pre-generative-AI baseline; comparable historical coverage of posted prices is much thinner elsewhere.

We treat Fiverr as the measuring instrument for a price that buyers and sellers on both Fiverr and Upwork care about. Extending coverage to other platforms is future work.

11. How is the index calculated? (the formulas)

The IPI is a matched-model index built in three steps, following the approach the BLS uses for CPI items that are hard to quality-adjust. The site recomputes the composite live in your browser whenever you change the basket.

Step 1 — Price relatives (same gig, period to period)

For every gig i observed in two consecutive quarters, compute the ratio of its new price to its old price:

p_i,t = price of gig i in quarter t (median Basic price if a gig has several snapshots that quarter). Relatives outside the band 0.1–10× are dropped as data errors, and a category-quarter needs at least 3 matched gigs to count.

Step 2 — Category index (Jevons, chained)

Within each category, combine that quarter's price relatives using a Jevons index — the geometric mean of the relatives — and chain it onto the previous quarter's level:

I^c_t = price index for category c in quarter t.  S_c,t = the set of gigs in category c matched between t−1 and t; |S_c,t| is how many there are. The product-then-1/n-power is exactly the geometric mean of the relatives. The base quarter (2020 Q1) is fixed at 100.

Step 3 — Composite IPI (weighted geometric mean)

Combine the category indices into one headline number with a Törnqvist-style weighted geometric mean:

w_c = the weight of category c (see below). Taking the weighted average in logs and then exponentiating is what makes this a geometric — rather than arithmetic — mean. Only categories you have selected enter the sum, which is why the headline updates as you toggle the basket.

Step 4 — The headline change

The number at the top is the percentage change in the composite from the base quarter (2020 Q1, 0) to the latest quarter (T):

The same formula gives each category's Δ'20–'26 column, using that category's index in place of the composite.

12. Are these prices adjusted for general inflation?

Not yet — the index is in nominal US dollars. It tracks the actual posted price of a gig over time, with no deflation by CPI or any other inflation measure. So a positive IPI reading does not automatically mean the work got more expensive in real terms — part of any rise can simply be economy-wide inflation over the same window.

For the AI-substitution story, what is most telling is when gig prices fall (or rise more slowly than general prices) despite a backdrop of broad inflation. A real (inflation-adjusted) version is a planned addition; until then, read the headline as a nominal change and keep the macro backdrop in mind.

13. How are the category weights set?

Like CPI expenditure weights, IPI weights are meant to reflect how much economic activity each category represents. We proxy transaction volume with review counts (a gig accumulates reviews roughly in proportion to sales). For each category we sum the maximum observed review count of its gigs, then normalize so the weights add to 1:

R_c = total review volume of category c. In the current sample design dominates the basket (~71%), with writing (~11%) next and marketing, coding, video, audio, and translation making up the rest. You can see each category's weight in the Weight column on the index page.

14. Why geometric means instead of plain averages?

Two reasons, both standard in official price statistics:

• Symmetry. A price that doubles and then halves should net to no change. The geometric mean (Jevons) gives exactly that; an arithmetic mean of relatives would show a spurious increase.
• It is the BLS standard. The geometric mean is what the Bureau of Labor Statistics uses for elementary CPI aggregates, so the IPI stays comparable to how real inflation is measured.

15. Could a few sellers distort the index?

The design is fairly resistant to a handful of outliers. Three guards do most of the work:

• Matched pairs only. A gig's price is compared to its own past, so new or vanishing sellers cannot swing a level.
• Geometric mean. It damps extreme relatives far more than an arithmetic mean would.
• Guardrails. We drop relatives outside 0.1–10× as data errors and require ≥3 matched gigs before a category-quarter counts.

No single seller's price change can move a category much, and the composite further averages across categories. The bigger risk is thin coverage (too few matches), which we flag in the limitations, not manipulation by any one seller.

16. Why can I switch categories on and off?

The composite is recomputed in your browser from the category indices and weights every time you change the basket — using exactly the Step 3 formula above, renormalized over whatever categories are selected. This lets you ask "what does the index look like for just design and writing?" without trusting a server to do it. Toggle with the checkboxes, or the All / None links.

17. Can I look up individual freelancers or a specific gig?

Yes. Beyond the composite chart, the site lets you drill into the top freelancers within each category and open an individual gig to see its own posted price over time.

This is the raw material the index is built from — a matched gig is just one seller's price compared to its own earlier price — so the per-gig view is the most direct way to sanity-check what the index is summarizing. It is also the quickest way to see why thinly-covered categories can look flat: with few gigs archived per quarter, there simply are not many lines to move.

18. What are the limitations and caveats?

• Pilot scale. These figures come from a sample of sellers, not the full marketplace. Treat them as indicative, not definitive.
• Thin categories read flat. Sparse matched-pair coverage can sit at exactly 100 for long stretches — that reflects missing matches, not genuine price stability. The earliest quarters and the smallest categories (translation, audio) are most affected.
• Design dominates. Because design carries ~71% of the weight, the composite mostly follows design. Toggle it off to see the rest of the basket on its own.
• Posted, not transacted. We observe list prices on the Basic tier, not final amounts paid after add-ons, discounts, or negotiation.
• Window starts 2020. The pre-2020 archive is too thin to chart, and the opening quarters overlap the COVID shock.
• Survivorship and archiving gaps. The Wayback Machine does not snapshot every page every quarter, and gigs that disappear leave the panel.
• Association, not proven causation. The index documents how prices move; attributing those moves to AI specifically requires the further analysis in the paper.

19. Can I reproduce this?

Yes. The index series shown here come from the project's analysis pipeline in code/ — the full-history quarterly build is code/18-build-site-data-long.py, which serializes to docs/data.json; the page reads that file and recomputes the composite client-side. The data contract and build steps are documented in README.md and GUIDE.md alongside this page.

20. How often is the index updated?

The series is rebuilt from the archive rather than streamed live, so it updates when we re-run the pipeline against fresh Wayback Machine snapshots — not continuously. Each build stamps the page with a generation date and the quarters it covers, so you can always see how current the shown series is.

Because the newest quarter depends on pages that have actually been archived and matched, the most recent point can shift slightly as more snapshots land, and firms up as that quarter fills in.

21. Found an error, or want to contribute?

The project is open. Code, data-build scripts, and this page live at github.com/AISmithLab/IntelligencePriceIndex. If you spot a misread price, a misclassified gig, or a bug in the pipeline, please open an issue or a pull request there. Methodology suggestions are welcome too — the index is meant to be auditable, and corrections from people who actually price this work on Fiverr and Upwork make it better.