Back to Blog
Research

Why TikTok Search Results Keep Changing (We Measured It)

November 22, 2026
12 min read
S
By SociaVault Team
TikTokTikTok SearchTikTok APIResearchTrend Tracking

TikTok search results change because each request gets a freshly assembled page instead of a copy of one fixed ranking. When we ran the same keyword search four times over eight minutes, at most 10 of the 30 videos showed up in all four results, and a single search surfaced between 59 and 87 different videos.

That matters if you track where a video "ranks" for a keyword, or compare two snapshots and call the difference a trend. Most of that difference is churn. Here's what we measured, what stayed put, and how to sample search so the numbers mean something.

What we ran

On 17 November 2026, between 16:00 and 16:38 UTC, we called our TikTok endpoints as plain logged-out requests:

  • keyword search for "space", "cooking" and "skincare", with no region set
  • the same "space" and "cooking" searches with region=US, run right after
  • hashtag search for #space, with region=US
  • the US trending feed
  • as a control, the first page of comments on one NASA video

Each request ran four times, about 2 minutes 45 seconds apart, and we only looked at the first page. That's 32 requests in total, at 1 credit each.

The results

RequestResults per callKept from the previous callIn all 4 callsDifferent results seen
"space"3012 to 17761
"cooking"309 to 19572
"skincare"3012 to 211059
"space", region US308 to 23563
"cooking", region US29 to 300 to 14087
#space, region US18 to 2011 to 171029
US trending feed13 to 160053
Comments on one NASA video15151515

Keyword search keeps about half its page

Without a region, back-to-back calls shared 15 of their 30 videos on average, anywhere from 9 to 21. Over four calls, each keyword turned up about twice as many videos as fit on a page. The top of the page wasn't much steadier: back-to-back calls shared between 3 and 9 of their top 10.

The number one spot held best. The same video led all four "skincare" calls, three of four for "space" and two of four for "cooking".

Videos that stayed didn't keep their place, either. A video showed up in two back-to-back calls 139 times across the three searches, and only 17 of those times was it in the same position. The usual move was 2 or 3 places.

Big videos stick, the long tail rotates

In all five keyword runs, the videos that appeared most often had the highest median play count. For "cooking", the ones in all four calls had a median of 3.0 million plays, against 1.2 million for videos we saw only once. For "skincare" it was 346,000 against 88,000.

So the page looks like a core of strong performers plus a rotating set of newer or smaller videos. We can't see TikTok's ranking, so take that as a description of what came back. It doesn't explain why.

Setting a region didn't steady anything

We expected region=US to tighten things up. It didn't. "space" kept 5 videos across all four calls instead of 7, and "cooking" kept none.

The "cooking" result is the strange one. One call came back with an almost entirely different page: 29 of its 30 videos weren't in the call before it, and they were much newer, with a median age of about 14 hours against 2 to 11 days in the other calls. The next call went back to the usual mix. If you'd taken that one snapshot, you'd have reported a dramatic shift in what ranks for "cooking" that was gone by the next call.

The region didn't change who showed up much either. Around four in five "cooking" videos came from US accounts with or without it. Use region when you care about a market. Just don't expect it to make results repeatable.

Hashtag pages have a stable core

The #space hashtag feed was the steadiest search we ran. Ten of its roughly 19 videos showed up in all four calls, and they were old, with a median age of 354 days. The videos we saw only once were a median of 2.3 days old. A hashtag page looks like a set of established videos with a few fresh slots that rotate.

Not one video appeared in two back-to-back trending calls. Four calls returned 53 different videos in 60 slots, and the 7 repeats only came back two calls later. A single call is a handful of what's being pushed in that region at that moment, and nothing more.

Comments didn't move at all

The control behaved. All four calls returned the same 15 comments in the same order. Only the numbers moved: the top comment went from 5,359 likes to 5,357, and the video's comment count went from 1,956 to 1,960. That's a useful baseline. Where TikTok's own data is stable, the responses are too, so the churn in search and trending is coming from TikTok.

Why it happens (our best guess)

This part is inference. We can see what came back. We can't see how TikTok chose it.

  • New uploads get tried out. The videos that rotated were newer and smaller than the ones that stayed, which is what you'd see if fresh videos get short trial runs in search.
  • Each request is a new visitor. A logged-out request has no watch history, so nothing ties one call's page to the next.
  • Some calls seem to get a different ranking altogether. The one-off "cooking" page full of hours-old videos didn't look like a reshuffle. It looked like a different version of the results.

We only tested logged-out requests. Logged-in users have their watch history in the mix as well, and we didn't measure what that adds.

How to track TikTok search without fooling yourself

  • Sample, then count appearances. Run the same search 3 to 5 times across the period you care about. Score each video by the share of samples it appeared in and its median position. A video in 4 of 4 samples is part of the core. One that showed up once may just have been passing through.
  • Track the share over time. "Our video was in 3 of 4 samples this week and 1 of 4 last week" is a real signal. "We dropped from 4th to 9th" usually isn't.
  • Dedupe by aweme_id. Across samples and across pages. Each page is a fresh request against a moving list, so page 2 can repeat something from page 1.
  • Collect trending in bulk. One trending call tells you very little. Fifty spread across a day show you which videos, sounds and hashtags keep coming back.
  • Find videos with search, then measure them elsewhere. Once you have the IDs, track their numbers through video-info or the creator's video list, which return the same video every time you ask.

Code: sample a search and score the videos

This runs a keyword search a few times, collects every video it sees with its positions, and ranks them by how often they appeared. It caps the number of samples so a typo can't burn your credits, and it handles list fields that arrive as objects keyed "0", "1" and so on.

// Node 18+. Run: SV_API_KEY=sk_live_... node sample-search.mjs
const BASE = "https://api.sociavault.com/v1";
const headers = { "x-api-key": process.env.SV_API_KEY };
const asList = (v) =>
  Array.isArray(v) ? v : v && typeof v === "object" ? Object.values(v) : [];
const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
const median = (xs) => {
  const s = [...xs].sort((a, b) => a - b);
  const m = Math.floor(s.length / 2);
  return s.length % 2 ? s[m] : (s[m - 1] + s[m]) / 2;
};

async function sampleSearch(
  query,
  { samples = 4, gapMs = 150_000, region } = {},
) {
  const seen = new Map();
  let ok = 0;
  for (let i = 0; i < Math.min(samples, 10); i++) {
    if (i > 0) await sleep(gapMs);
    const params = new URLSearchParams({
      query,
      ...(region ? { region } : {}),
    });
    const res = await fetch(`${BASE}/scrape/tiktok/search/keyword?${params}`, {
      headers,
    });
    if (!res.ok) {
      console.warn(`sample ${i + 1}: HTTP ${res.status}`);
      continue;
    }
    ok++;
    const { data } = await res.json();
    asList(data?.search_item_list).forEach((item, index) => {
      const v = item?.aweme_info;
      if (!v?.aweme_id) return;
      const row = seen.get(v.aweme_id) ?? {
        id: v.aweme_id,
        author: v.author?.unique_id,
        positions: [],
      };
      row.positions.push(index + 1);
      row.plays = v.statistics?.play_count ?? row.plays;
      seen.set(v.aweme_id, row);
    });
  }
  return [...seen.values()]
    .map((r) => ({
      ...r,
      share: r.positions.length / ok,
      medianPosition: median(r.positions),
    }))
    .sort((a, b) => b.share - a.share || a.medianPosition - b.medianPosition);
}

const rows = await sampleSearch("skincare", { samples: 4 });
console.table(
  rows.slice(0, 15).map(({ id, author, share, medianPosition, plays }) => ({
    id,
    author,
    share,
    medianPosition,
    plays,
  })),
);

And in Python:

# Run: SV_API_KEY=sk_live_... python sample_search.py
import os, statistics, time
import requests

BASE = "https://api.sociavault.com/v1"
HEADERS = {"x-api-key": os.environ["SV_API_KEY"]}

def as_list(v):
    if isinstance(v, list):
        return v
    return list(v.values()) if isinstance(v, dict) else []

def sample_search(query, samples=4, gap_s=150, region=None):
    seen, ok = {}, 0
    for i in range(min(samples, 10)):
        if i:
            time.sleep(gap_s)
        params = {"query": query, **({"region": region} if region else {})}
        r = requests.get(f"{BASE}/scrape/tiktok/search/keyword", headers=HEADERS, params=params, timeout=90)
        if r.status_code != 200:
            print(f"sample {i + 1}: HTTP {r.status_code}")
            continue
        ok += 1
        data = r.json().get("data") or {}
        for index, item in enumerate(as_list(data.get("search_item_list"))):
            v = (item or {}).get("aweme_info") or {}
            if not v.get("aweme_id"):
                continue
            row = seen.setdefault(v["aweme_id"], {"id": v["aweme_id"], "author": (v.get("author") or {}).get("unique_id"), "positions": []})
            row["positions"].append(index + 1)
            row["plays"] = (v.get("statistics") or {}).get("play_count", row.get("plays"))
    rows = [{**r, "share": len(r["positions"]) / ok, "median_position": statistics.median(r["positions"])} for r in seen.values()] if ok else []
    return sorted(rows, key=lambda r: (-r["share"], r["median_position"]))

for r in sample_search("skincare", samples=4)[:15]:
    print(f'{r["share"]:.2f}  pos {r["median_position"]:>4}  {r["plays"] or 0:>10}  @{r["author"]}  {r["id"]}')

Each sample is one request and costs 1 credit, so four samples of one keyword cost 4. The default gap of two and a half minutes matches our test. To track a keyword over a week, spread the samples across days instead and keep each day's results, so you can compare shares from one week to the next.

What this test can't tell you

  • It's one afternoon. Four calls per search over about eight minutes, on one day. Churn could be higher at peak hours or lower overnight. We haven't measured either.
  • It's three topics. All broad ones. A narrow query with fewer matching videos has less to rotate in and may hold steadier. We didn't test that.
  • It's page one. Deeper pages might behave differently.
  • It's logged-out only. We can't speak for what a logged-in user sees.
  • It can't say why. The causes above are our reading of the data. We have no inside knowledge of TikTok's ranking.

If you're testing a TikTok data provider, this is also why two runs of the same search won't match, whoever you use. Judge a search endpoint on relevance and filters, as in our checklist for vetting a scraping API.

Frequently Asked Questions

Why do my TikTok search results keep changing?

Because TikTok builds a fresh page of results for each request instead of serving one fixed ranking. In our test, back-to-back keyword searches a few minutes apart shared about half of their 30 videos, and no more than 10 appeared in all four calls.

Does TikTok search show everyone the same results?

No. Even identical logged-out requests a few minutes apart got different pages in our test. Logged-in users also have their own watch history in the mix, which we didn't measure.

Don't track a single position. Run the search several times over the period you care about, then record the share of samples that included your video and its median position. A change in that share is a real signal. One rank on its own is mostly noise.

Each call returns a new handful of what's being distributed in that region. Four US trending calls a few minutes apart had no videos in common between back-to-back calls, and 53 different videos in total.

Does setting a region make TikTok search results stable?

Not in our test. With region=US, "space" changed just as much between calls as it did without a region, and "cooking" changed more, because one call returned a page that was almost entirely new. Use the region parameter to target a market, not to make results repeatable.

They were in our test. Ten of the roughly 19 #space videos showed up in all four calls, and those were mostly established videos close to a year old. The newer ones rotated.


Want to see how much your own keywords churn? Get 50 free credits, no card needed, and run the sampler above with four samples on a couple of searches. If you're using search for content ideas, our TikTok keyword research guide covers filters and sorting, and the TikTok search API overview lists every search endpoint.

Found this helpful?

Share it with others who might benefit

Ready to Try SociaVault?

Start extracting social media data with our powerful API. No credit card required.