AI recommendations shifted 6.1 points in 6 days, without brands changing anything
Citation-level evidence of a Perplexity retrieval update and what it means for your brand
A brand can disappear from AI recommendations entirely for one week and fully reappear the next. Between our March 20-21 and March 26 scans, 6 days apart, hiking boot brand scores shifted an average of 6.1 points with no brand-side content changes. We traced the cause to a Perplexity retrieval reweighting using citation domain data.
Disclosure: Envoyra sells the AI visibility scanning tool used to collect this data. We have a financial interest in findings that demonstrate the value of ongoing monitoring. The methodology is disclosed in full below. Underlying scan data, including citation domain frequencies, is available to clients and press on request.
Background: four weeks of weekly scans
We have been running weekly scans of 25 hiking boot brands on Perplexity Sonar since early March 2026. Each scan fires the same 60 buyer-intent prompts against the same brand set using the same methodology. Four scans have been completed to date:
Scan 1 (early March): Initial baseline. Partial coverage, several brands entered the dataset for the first time. Scan 2 (March 13-14): First full-coverage scan across all 25 brands. Scan 3 (March 20-21): Second full-coverage scan. Scan 4 (March 26): Third full-coverage scan.
This article compares Scan 3 (March 20-21) and Scan 4 (March 26), the two most recent scans with complete cross-brand data. The 5-6 day gap between them is the shortest interval in our dataset, making it the cleanest test of whether score movement reflects brand-side changes or platform-side retrieval shifts.
The finding in one sentence
A brand can disappear from AI recommendations entirely for one week and fully reappear the next, with no changes to their website, their content, or their marketing.
Between our March 20-21 scan and our March 26 scan, 7 of 25 hiking boot brands gained ground and 7 lost ground on Perplexity Sonar. None of them changed anything. The movement was caused by a shift in which sources Perplexity pulls from, and we have the citation data to show it.
Dataset summary
Platform: Perplexity Sonar (standard tier). Vertical: Hiking boots. Brands tracked: 25. Prompts per scan: 60 buyer-intent queries. Scan 3 date: March 20-21, 2026. Scan 4 date: March 26, 2026. Gap between scans: 5-6 days. March 20-21 total citations analyzed: 2,196. March 26 total citations analyzed: 2,218.
Prompts are identical across scans: all scripts query the same DynamoDB table using the same GSI with the same ACTIVE filter. Same 60 prompt IDs, same text, same sort order. No prompt was added, removed, or edited between scans.
What changed in brand scores
Between March 20-21 and March 26, 7 brands rose, 7 fell, and 11 were approximately stable, without any observed brand-side content changes.
Salomon: March 20-21 score 70.0 → March 26 score 78.3 (+8.3). Oboz: 32.8 → 45.0 (+12.2). LOWA: 31.5 → 38.3 (+6.8). KEEN: 40.7 → 45.0 (+4.3). Hoka: 46.3 → 50.0 (+3.7). Merrell: 74.6 → 75.0 (+0.5). Columbia: 28.3 → 28.3 (0.0). La Sportiva: 41.8 → 35.0 (-6.8). Zamberlan: 24.1 → 16.7 (-7.4). Scarpa: 22.2 → 16.7 (-5.5). Danner: 20.0 → 16.7 (-3.3).
Category average absolute movement: 6.1 percentage points per week. Category median absolute movement: 3.5 percentage points per week.
Why we concluded this was model drift, not content change
Four lines of evidence point to a Perplexity retrieval update rather than brand-side content changes:
1. Movement is symmetric, not isolated. If a brand changed its content, only that brand's score would move. Instead, 7 brands rose and 7 fell in the same week. Coordinated reshuffling across the category is a retrieval event, not a content event.
2. Mention density stayed flat. March 20-21: 262 total mentions across 1,372 responses = 0.191 mentions per response. March 26: 289 total mentions across an estimated comparable response count = 0.193 mentions per response. Perplexity is not recommending more brands per response. It is recommending different brands. This is substitution, not expansion.
3. Salomon shows classic mean reversion. Salomon's inclusion rate across three consecutive weekly scans: 79.8% → 71.7% → 78.3%. This is a brand oscillating around a stable equilibrium, not a brand that improved its content, lost ground, then improved again in three weeks. Content-driven changes do not produce this pattern.
4. The winners and losers cluster by source type. Brands that gained (Oboz, KEEN, LOWA, Salomon) all have deep coverage on US-based structured gear comparison sites. Brands that lost (Zamberlan, La Sportiva, Scarpa) are European specialty brands with thinner English-language review site coverage. This pattern is not consistent with individual content changes. It is consistent with a retrieval source reweighting.
The citation evidence
We pulled citation URLs from S3 for both scans and compared domain frequencies. Total citations: 2,196 (March 20-21) vs 2,218 (March 26), nearly identical, confirming source substitution not expansion.
Domains that gained citations between scans: cleverhiker.com (+16, from 97 to 113). myoutdoorbasecamp.com (+15, from 3 to 18). backpackinglight.com (+14, from 20 to 34). outdoorgearlab.com (+14, from 207 to 221). meindlusa.com (+8, new entry).
Domains that lost citations between scans: switchbacktravel.com (-28, from 113 to 85). mountaineerjourney.com (-18, from 188 to 170). hikingfeet.com (-11, from 67 to 56). bakershoe.com (-10, from 10 to 0). rei.com (-9, from 214 to 205). treelinereview.com (-8, from 71 to 63).
What this reveals: The shift is from blog-style review sites (switchbacktravel -28, mountaineerjourney -18, hikingfeet -11) toward structured gear comparison and database-style sites (outdoorgearlab +14, cleverhiker +16, backpackinglight +14).
This is not the pattern we originally hypothesized. Our initial read suggested REI gained weight. The data shows REI actually dropped 9 citations. The real shift is more specific: Perplexity deprioritized opinionated review blogs and boosted structured gear databases between these two scans.
Why Oboz gained the most
Oboz's +12.2 percentage point jump is the strongest evidence for a retrieval event rather than a content change.
Oboz did not update their website between March 20-21 and March 26. What Oboz does have is extensive coverage on OutdoorGearLab and CleverHiker, two of the three biggest citation gainers in the March 26 data. When Perplexity increased the retrieval weight of those sources, Oboz was the brand with the most to gain from that specific shift.
This is the clearest demonstration in our dataset that AI recommendation scores are partially determined by factors entirely outside a brand's control. A brand can have strong structural content and still lose ground if the review sources covering them fall out of favor with the retrieval index. Conversely, a brand with deep third-party coverage can gain ground without doing anything.
If this is happening in your category, it's, and you'd never know without measuring. Explore the research →
The volatility index
Across four weekly scans (early March through March 26), here is the measured volatility in this category on Perplexity Sonar:
Raw (all transitions): 39 transitions measured. Mean absolute change: 5.6pp. Median absolute change: 2.0pp. Transitions >5pp: 13. Transitions >10pp: 4. Max single move observed: 46.3pp.
Adjusted (excluding first-scan entries): 24 transitions measured. Mean absolute change: 5.3pp. Median absolute change: 2.8pp. Transitions >5pp: 9. Transitions >10pp: 2. Max single move observed: 41.8pp.
Adjusted for dataset initialization, the average weekly score movement across 25 hiking boot brands on Perplexity Sonar is 5.3 percentage points, measured across 24 week-over-week transitions from early March to March 26, 2026.
On the Hoka 46.3pp figure: Hoka had no data in the first scan (first entry). Its initial swing reflects entering the dataset from zero, not movement from an established baseline. Excluded from the adjusted set.
On La Sportiva: La Sportiva first appeared in our earliest scan at 25.0%, dropped to 0.0% the following week, then jumped to 38.3% in the March 13-14 scan, a +41.8pp move from a genuine zero, not an initialization artifact. This is real volatility: a brand can disappear from AI recommendations entirely for one week and fully reappear the next. Included in the adjusted set.
The most stable brand observed: Columbia at 0.8pp average movement across all transitions, nearly flat across four weeks. The contrast between Columbia (0.8pp) and La Sportiva (41.8pp peak swing) captures the full range of volatility this category exhibits.
What this means for measurement methodology
You cannot measure AI visibility with a single snapshot. A brand that scores 45/100 on a single scan could be at its floor, its ceiling, or its equilibrium point. Without weekly tracking, you cannot tell.
Score changes require causal attribution before action. A brand that sees its score drop 6 points this week should not immediately rebuild its website. The first question is whether the drop is model drift (external, no action needed) or a competitor's content improvement (external, strategic response needed) or a signal gap in their own content (internal, fixable).
Citation-level data is necessary for attribution. Score-level data shows that something changed. Citation-level data shows why. Without the domain frequency comparison in this analysis, we would have had a plausible hypothesis (Oboz improved their content) that was factually wrong.
Temperature pinning matters for methodology integrity. Our scan payload did not explicitly set temperature for these scans. We relied on Perplexity's server-side default. We have since added explicit temperature pinning (temperature: 0.2) to eliminate that source of variance from future scans. Score changes from late March onward will be isolated from temperature variance.
Limitations
This analysis covers one platform (Perplexity Sonar), one vertical (hiking boots), and a 5-6 day gap between scans. We cannot confirm whether the same retrieval shift affected other verticals or other Perplexity models. We inferred the retrieval update from citation domain frequency changes. Perplexity has not publicly announced any model or index update during this period. The causal chain (retrieval reweighting → domain shift → brand score movement) is our best explanation of the data, not a confirmed mechanism.
What this means for you
If you are measuring AI visibility quarterly, you are blind. If you are measuring it monthly, you are late. If you are measuring it weekly, you can separate three distinct types of movement:
Model drift: Perplexity reweighted a source. External. No action needed. Competitor gains: a competitor improved their content or citations. External. Strategic response needed. Structural gap: your own signal architecture has a fixable weakness. Internal. Act on it.
Without weekly tracking and citation-level data, all three look identical. A brand that sees its score drop 6 points and immediately rebuilds its website may be solving the wrong problem entirely.
We monitor this weekly: score movement, citation domain shifts, competitor displacement patterns, and the three highest-leverage structural fixes available at that moment.
Explore the research
Most brands assume they show up in AI recommendations. Most don't.
The research documents which brands appear, which are consistently absent, and what the structural differences are.
Data: March 20-26, 2026. 60 prompts. 25 brands. Perplexity Sonar. Citation analysis: 2,196 (March 20-21) + 2,218 (March 26) citations from S3 stored response data. Temperature not explicitly pinned in these scans. Fixed in late March+. Score changes not attributable to temperature variance based on symmetric movement pattern. Envoyra Engine v1.0. envoyra.com. March 2026. Built by enterprise technology advisors with backgrounds in global digital strategy and technology sales.