When you open Pinterest, an invisible mechanism decides in a fraction of a second what appears on screen. Here is how it reads you, and how to take advantage of it.
Every pin on your home feed arrives with a reason attached: the name of the signal that selected it for you. These names exist for real, identical on the app and on the website. By reading them, we reconstruct the way Pinterest understands you. Three acts: what it knows about you, how the machine works, and how to take back control.
Before explaining the machine, let's look at what it already knows about you. It's more precise than you think, and it's written in plain text.
Every pin on your home feed carries a secret label that says why it is there.
This word is not decorative. It's the signal that triggered this pin's display. There are 263 of them, one for each way of guessing what might catch your eye.
Pinterest doesn't just guess your profile: it writes it out, in black and white, in the page your browser receives.
Age range, gender, region, interests bought by advertisers, and even the fact that an advertiser you've already visited « follows » you: it all travels in plain text. Want to see yours? The Mirror module decodes it for you.
Before it even classifies it, the algorithm « reads » your visual with several AI models. Here is what it extracts, measured on our captures.
Pinterest writes its own description of your image (« a white plate with a slice of bread… »). Present on 77 of 80 pins.
auto_alt_textThe text written INSIDE the image is read and indexed, just like the title. Your keyword should appear there.
OCRDetection boxes locate the objects in the visual (x, y, width, height zones).
visual_objectsThe main hue is extracted from every pin. Present on 80 of 80 pins.
dominant_colorA model ties the pin to an interest (« food and drinks »), without relying on your tags.
gen_ai_topicsA unique perceptual signature. It's the « seed » used to suggest neighbouring pins to you.
image_signatureDirect takeaway for your pins : a sharp visual, a keyword written in the image, a filled-in alt text. Pinterest doesn't just store your image, it understands it.
Millions of pins go in one end, twenty-five come out the other. In between: a three-stage pipeline and a gallery of specialised signals.
Select, score, assemble. The final score, though, never leaves Pinterest's servers.
Millions of pins narrowed down to a few thousand. This is where the 263 signals act.
A model scores every pin. That score decides the order, and stays on the server, never shown.
Blending, diversity, and ads. On the web, about 4 of the first 10 pins shown are sponsored, often right at the top of the page.
Real composition of a top 10 (desktop web) : 4 sponsored pins, at positions 1, 3, 4 and 9.
The 263 signals split into 9 headline reco families, plus 7 cross-cutting or technical families (freshness, shopping, formats, ads, onboarding, system). Here are the 9 main engines first; all 16 families are explorable below.
Re-shows what already worked on you, or on people like you.
« People who saved this also liked… »
A model of you, broken down into several interests.
Reads the pin's image and text to find its neighbours.
Guesses your intentions and builds recommendations for you. It also drives your notifications (the NOTIF_* families).
What you type into search also feeds your home feed.
Quotas for freshness, creators, categories. The editor's invisible hand.
A digital signature of each pin's text and video.
Brands visited, browsing elsewhere: it all flows back here.
« Gemini V5 » is an internal Pinterest model, unrelated to Google Gemini.
The full list, exactly as it appears in the code: 9 reco families, 7 cross-cutting families, 263 signals in total. Filter by family. Format: code SIGNAL_NAME.
Depending on the signal that selects you, the content doesn't carry the same popularity. We measure it by the number of saves.
The families above map to a decade of Pinterest publications:
Sifter and Intentional Distribution, on the other hand, have no public paper: that's the real blind spot.
For the PINNABILITY_CONDITIONAL_PIN_EMBEDDINGS sources, a pin's "reason" points to a seed pin: a pin you saw or engaged with recently. Decoding the tracking of a real save, we read f6=232 (the source) and f8=2203789a2be7b55b2fa6008c8aedfbcf. That hash is not the displayed pin's image: it's the image_signature (visual fingerprint) of another pin, the seed. In other words, the candidate cluster is addressed by the visual content of a recent pin. "Because you saved something similar" is literal.
Your action doesn't vanish into the void. It comes back to reinforce the signal that showed you the pin.
You save a pin
The app tells Pinterest which signal had selected it
That signal is reinforced, for you, within seconds
A save is worth more than a like. Time spent counts as much as the click. And engagement that lasts months weighs more than a one-day spike.
Impression
seen, and logged
Outbound click ≥ 5s
the "good click", Pinterest's take on NavBoost's longest good click
Click ≥ 120s
strong engagement, memorised
Pinterest counts a qualified outbound click from 5 seconds (counter AUTOMAGICAL_BOARD_PIN_5S_OUTBOUND_CLICKS): the direct equivalent of NavBoost's "longest good click". A second threshold, at 120 seconds of outbound click, triggers the local memorisation of an interest. Dwell time is measured continuously through TIMED_PAIR event pairs: in our live captures, from 0.35s to 181.8s on a single pin.
We see which signal selected the pin. We never see the score that decides the order.
Feed pagination does not use a plain page number. The cursor (the "bookmark") encodes a score boundary: an INT64_MAX − N value where N derives from the ranking score. To serve the next page, the server resumes "below" that boundary. That's the signature of an ordering driven by a continuous server-side score: if the client ordered the feed itself, it wouldn't need this score cursor.
NavBoost is the exact name of a Google system, made famous by the 2023-2024 antitrust trial. Yet Pinterest uses the same word in its own code, and had even published it as early as 2018: a signal that measures clicks, long clicks and saves, segmented by country, gender and time window (7 days, 90 days, 2 years). Coincidence? Or a Google defector who moved to Pinterest with the concept in their luggage?
The wink repeats: Pinterest also has an in-house component called Gemini (GEMINI_V5_*), which has nothing to do with Google's Gemini. Borrowing the names that make headlines is an internal habit. Seven years on, NAVBOOST_* signals still surface your pins.
Source: Pinterest, « Demystifying Core Ranking in Pinterest Image Search », arXiv:1803.09799 (2018)
Understanding the machine is good. Making it work for you is better. On to the tests and the concrete levers.
The algorithm reacts almost in real time, and it forgets just as fast.
Five « bonsai » pins saved, and the home feed switches to plants within seconds.
A single « Japanese bonsai » search and botanical content slips into the home feed, with no visible « search » label.
One click on « see less » and the source disappears. Reversible: the algorithm forgets as fast as it learns.
What our tests showed, Pinterest now confirms publicly: since 2026, recommendations update while you browse, instead of waiting for a retraining.
Two sources: our analysis and the 2026 public watch.
Without history, the content-understanding signals are what surface you. Your only lever: a distinctive visual, a clear title and description.
This is the number-one real-time lever. A quick save triggers redistribution within seconds.
« Because you saved something similar » starts from the visual fingerprint of a pin seen recently. A consistent visual places you in the right clusters.
The feed de-duplicates visually look-alike pins (the near_dup_sig signature). Posting ten near-identical variants cannibalises them: clearly distinct visuals work better.
Pinterest reads the text inside the visual (OCR). Your main keyword should appear there, not just in the title.
Always fill in the alt text: Pinterest indexes it just like the title. Community analyses link it to a clear rise in impressions and outbound clicks.
The board where you first save a fresh pin shapes the understanding of your content. Name it precisely.
Consistently covering a topic strengthens where you're filed in the right interest: this is what the PinnerSage and topic signals drive (RECOMMENDED_TOPICS, FOLLOWED_INTEREST, the top_interest field attached to each pin). Alignment matters more than volume.
Pinterest Assistant answers multimodal conversational queries. Structure your content as an answer to an intent, situational and seasonal.
A missed impression is never final. Freshness is a lever: publish regularly rather than in bursts.
Paste the source code of your Pinterest home feed. The tool hands you back what Pinterest knows about you, and why each pin is there.
The score that decides your feed's order (pinnability_multi_scores) is absent from the page. A number governs your home feed, and it appears nowhere here.
