The protein craze is cooling. The protein habit isn't.

The question

Singapore's search interest in protein peaked in April 2026 and has fallen about a third since, but it is still well above where it was a year ago. The spike was hype. The habit underneath it is real, and it is still growing.

I started this because of a headline. On 8 May 2026, CNA ran a Big Read asking whether the high protein craze is a fad or a fixture in the Singaporean diet. It made its case with interviews, brand quotes and market figures. What it didn't have was a long view of public interest over time, so I went and built one from Google Trends, and used it to test the article's claims one by one.

This is a personal project. Nobody asked for it; I wanted to know.

The data

I used Google Trends for Singapore: monthly search interest from 2004 to October 2026, scored 0 to 100. I downloaded every comparison by hand, because the usual Python library for Trends is archived and gets blocked, and the official API is still invite only.

Google scales each download 0 to 100 on its own, so two separate files can't be compared directly. My workaround was to put Protein (the Nutrient topic, not the plain word) into every download as an anchor. Where the Protein column came out identical across files, I knew every other term sat on the same scale and could be merged.

What this data can and can't tell you:

• It measures attention, not eating or buying. A rise in searches is people thinking about protein, not proof they eat more of it.
• It has no age or gender breakdown, so questions like "is this a women's trend or a men's trend" are out of reach.
• Small terms are fragile. A topic that moves between 3 and 6 is mostly rounding noise, so I only drew conclusions from direction, not decimals, for anything that small.
• I start the analysis in January 2019. Before that, the Protein topic has a data problem (see What almost fooled me).

What I found, in nine questions

Chart: yearly averages of Singapore search interest, 2019 to October 2026. See it live in the interactive story.

I worked through the project as a chain of questions, each one raised by the answer before it. Every question below has an interactive chart in the scroll story, where you can draw your own guess for the first one.

1. Is the protein hype already over?

Not over. Cooling. Down 37% from the peak, but still above the same month last year.

I put 12 month Bollinger bands on protein: a rolling average with a band two standard deviations either side. January, March and April 2026 all broke above the upper band, which is interest running beyond its own normal range. That is what hype looks like in data. Since April it has dropped from 100 to 63, but it has stayed inside the band, so this is a cool down rather than a collapse. October 2026 is still well above October 2025, which sat at 45.

One pattern to keep in mind while it cools: the season is a year end slump, not a January spike. Once I removed the trend, January was only 3% above a normal month, while December runs about 15% below normal, for keto as well as protein. So a dip this December won't, on its own, mean the craze is ending.

2. Will it crash like keto did?

Unlikely. Keto lost most of its searchers. Protein keeps landing higher.

Keto peaked at 35 in June 2020, during the circuit breaker, and scores 3 today. The better test was the floor. Every time protein's 3 month average dipped below its 12 month average, I noted where the 12 month average sat: 23.5, 26.8, 26.9, 25.2, then 42.4, then 71.0. Since 2023, each dip has landed higher than the one before. Keto went the other way: 24.2, 10.3, 5.3.

It's tempting to say protein took keto's place, since the lines cross and their levels correlate at minus 0.61. But when I compared changes instead of levels, month on month and year on year, they move together: +0.29 and +0.47. The timing doesn't fit a handover either. Keto halved between 2020 and 2022 while protein dipped too. Protein's surge came afterwards, from 2023, and it was far bigger than anything keto had to lose. That reads as new interest, not people switching diets.

3. Hype or habit?

Hype lit up early and faded. Protein lit up late and is still near its high.

Keto was one hype that burned out. To see whether protein looks like hype at all, I downloaded 32 food and fitness trends, scaled each to its own peak, smoothed them over 3 months and grouped them by shape with k-means clustering (k of 4, chosen after checking the elbow and silhouette scores). Four shapes came out:

• Burned out: keto, intermittent fasting, paleo, bubble tea, smoothie. A sharp peak around 2020, then a long slide.
• New surges: protein, whey protein, creatine, Hyrox, matcha, air fryer, Ozempic. Recent and still near their highs.
• Slow climbers: gym, calorie, high-protein diet, collagen, oat milk, pilates, meal prep and others. Steady growth over years.
• Steady staples: yoga, running, veganism, kombucha, CrossFit. Present the whole time.

Read as a whole, the four shapes are a life cycle. Burned out is what hype looks like: diets with strict rules that peaked around 2019 and 2020, mostly during the circuit breaker, then never came back. Steady staples are what a settled habit looks like: no big spike, no disappearance. Slow climbers are a habit forming: a little more every year. New surges are too early to call, and each will end up either burned out or a habit.

So which way is protein going? The word "protein" sits with the new surges, still riding the hype. But the high-protein diet, the actual eating habit, sits with the slow climbers and has been building since 2019. The craze is new. The habit underneath it is not. Many of the trends also peaked together in spring 2026, a clue I come back to in question 6.

The method showed its limit too. K-means put Dalgona coffee, a 2020 lockdown craze, with the new surges, because it groups by overall closeness, not by the story a curve tells. So I read the groups as a guide rather than a verdict.

4. Does searching mean doing?

Searches nearly doubled. Eating and training rose about a tenth.

Search only measures curiosity. My first attempt to check behaviour used SFA food supply data, and it failed: supply isn't eating (see the research log). So I went to the national surveys and the market figures instead, and lined up four links in the chain, 2020 to 2025:

Singaporeans are careful: reading about protein doesn't immediately turn into eating more of it. Buying moved more than eating, and the brand searches show where that money is going. Most protein powder brands spiked in March 2026 and dropped back (Dymatize to zero by October), while the ready-to-drink brands Rokeby and Meiji held close to their highs. People want convenience: ready-to-drink over powder.

Brand searches also expose a blind spot. Singaporeans often search the brand rather than the category. With oat milk, Oatside overtook Oatly in 2023 and now outscores it several times over, so a category search misses part of the interest. The same is likely true for protein, which means my protein line probably undercounts.

5. Who actually needs more protein?

Adults over 50, and the protein aisle isn't talking to them.

The National Nutrition Survey 2025 says 79% of residents already meet their protein needs. Split by age, it's 93% of adults aged 18 to 49 but only 55% of those aged 50 to 69. From 50, the body needs more protein per kilo to hold on to muscle, so the people who need it most are the ones protein marketing ignores.

There are early signs this group is moving. In the National Population Health Survey, strength training among 60 to 74 year olds rose from 25.5% in 2020 to 36.8% in 2025, and among women from 27.8% to 33.3%. MOH describes the overall level as stable, so I treat this as a hint rather than proof. It also fits the CNA claim that the trend is spreading to women in their 40s and 50s, which search data alone couldn't test.

6. Was it Hyrox?

No. It was a health wave, not a race.

Hyrox, the fitness race the CNA article points to, was the obvious suspect. It appears in Singapore searches in April 2023, just as protein's climb starts, and its biggest month, April 2026, landed on protein's peak.

To test it, I used causal impact: train a regression on other food trends before the event, use it to predict what protein should have done afterwards, and see whether the real line escapes the 95% normal range. The controls were matcha, air fryer, collagen and Ozempic searches, trained on November 2021 to March 2026.

Without fibre, April shoots above the range and Hyrox looks guilty. Add fibre and April falls inside it. Fibre jumped at the same time as protein, and nobody eats more fibre for a fitness race, so both were riding a wider health wave. Without fibre in the model, the only explanation it had was Hyrox.

The lesson I'll keep: in causal impact, the choice of controls is the answer. Hyrox still matters for the long climb since 2023. It just didn't cause the spike.

7. Could anyone have seen the turn coming?

No. The best model just copied last month.

I built lag features (the last three months, the same month last year and the calendar month), trained four forecasts on 2019 to 2025 and tested them on January to October 2026. Scored by mean absolute error, lower is better:

The random forest stayed flat near 49, because a random forest can't predict a value higher than anything in its training data, and the highest month it had seen was 58. The naive rule won, but only because it always arrives one month late. It never predicts a turn at all. A hype peak is exactly the kind of event that history-based models miss.

8. Is Singapore just following the US?

Not just the US. Asia is moving together, and Japan went first.

I first compared Singapore only with the US. The two tell the same story on a different clock: US protein interest first reached 70% of its own peak in January 2024, Singapore only in January 2026, and the US peaked in February 2026, two months before Singapore. US interest this October was 56 against 64 a year earlier, already below last year.

But the US alone couldn't tell me whether Singapore is copying America or riding something closer to home. So I added Japan, Hong Kong and Malaysia, using the Protein topic rather than the word, because Japanese searches aren't in English. Hong Kong and Malaysia climbed in step with Singapore through 2026, and like Singapore they are still above last year: Hong Kong at 45 against 34, Malaysia at 43 against 37. That makes it a regional wave, not just an American trend arriving late.

Japan is the most useful comparison, because it went first. In Japanese, プロテイン usually means protein powder, while タンパク質 is protein as a nutrient, and both come from one download, so they share a scale. The powder word climbed for years, peaked in June 2020, right after Japan's first state of emergency, and has never crashed: this October it sat at 56% of that peak, the same as a year earlier. Meanwhile the nutrient word kept climbing, from a yearly average of 23 in 2019 to 42 in 2026. In 2019, Japanese users searched タンパク質 44% as often as プロテイン; by 2025 it was 69%. The hype cooled and held, and the conversation moved from the shaker to the plate. That is the ending I expect for Singapore, and it fits what the market data says about Japan, where ready-to-drink protein is already a mainstream convenience store product.

Japan also shows how diet hype ends. Its version of keto was 糖質制限, carb restriction (ケトジェニック, the literal word for keto, barely registers). In the same download, on the same scale, carb restriction was searched more than タンパク質 in 2016. It peaked in February 2016 and has since fallen to 7% of that peak, a yearly average of 2.2 in 2026 against 42.1 for タンパク質. The two lines crossed in 2017 and never met again. That is the keto and protein story from question 2, played out in Japan years earlier: the diet with rules burned out, the nutrient stayed.

Korea tells the same story in its own language. It never had a powder craze: the loanword 프로틴 is rarely searched and has crept up slowly. The everyday word for protein as a nutrient, 단백질, rose from a yearly average of 14 in 2019 to 32 in 2026, and this October it is still above last year (32 against 25). No spike, no crash: the shape of a habit, not hype. The same downloads showed the English word "protein" jumping in November 2025 and collapsing by October 2026 in both Japan and Korea, in the same months as the Korean topic glitch, so I don't read anything into that line.

Keto tells the same story in reverse. It peaked at different times in each country (US January 2019, Malaysia and Singapore mid 2020, Hong Kong April 2021, Japan June 2022) and has faded everywhere, to between 11% and 18% of its peak in the three Asian markets.

Three caveats. Hong Kong and Malaysia both show a one-month spike in September 2026 that Singapore doesn't, and I couldn't find the cause. Japan's protein topic jumps from almost nothing to half its eventual peak in early 2021, which neither Japanese word shows, so I treat it as a change in how Google grouped searches and use the words instead. And I left South Korea's protein topic out: its protein and keto data both drop to near zero and jump back in the same months, which points to a Google data problem rather than people. The Korean word 단백질 doesn't have that problem, so Korea comes in through it instead.

9. So how long will it last?

The protein craze is cooling. The protein habit isn't.

My bet: interest keeps easing over the next 6 to 12 months, the way it already is in the US, but settles on a higher floor than before. Unlike keto, protein is a nutrient, not a diet, so it is harder to abandon. The habit is moving to the easiest format, ready-to-drink over powder. And the bigger opportunity isn't another protein bar for gym-goers. It's reaching the adults over 50 who genuinely need more.

Along the way I wrote down twelve hypotheses and tested each one. Of those, one was supported (people want convenience), two partly held (fitness, and Singapore following the US, which turned out to be a regional wave), and the rest were ruled out or weak: protein as the next keto, protein stealing keto's audience, weight loss, Hyrox causing the spike, Ozempic, searching as eating, a new year effect, fibre as the next protein, and a model that could predict it.

Testing the article's claims

The CNA piece came out on 8 May 2026, about a month after search interest peaked. Mainstream coverage tends to arrive at or just after the top of a hype cycle, and this one did. Here's how its main claims hold up against the search data (claims paraphrased).

The article relies on interviews and market sizes. What the search data adds is the shape over time, and that's what lets me separate the spike from the habit.

What almost fooled me

The findings above only hold because I caught these first. Each one would have given me a confident, wrong answer.

A glitch that looked like a crash. In November 2016, protein interest fell from 16 to 5 in a single month and stayed near zero for most of 2017. My first read was that protein "started" in 2018. Then I checked the other terms in the same months: keto rose, and calorie, low fat and cottage cheese held steady. Only the Protein topic dropped, so it was a change in how Google sorted searches into that topic, not a change in people. The US data later showed the same drop in the same month, which settled it. I start the analysis in 2019 because of it.

A trend dressed up as a season. My first seasonality chart showed keto falling smoothly from January to December, like a staircase. Real seasons go up and down. Keto was declining within every year, so dividing each month by its calendar year's average turned the trend into fake seasonality. Comparing each month with a centred 12 month average fixed it, and most of keto's "January peak" disappeared.

A correlation that pointed the wrong way. Protein and keto levels correlate at −0.61, which reads like replacement. Any two series trending in opposite directions will do that. Comparing changes instead flipped the sign, and the replacement story fell apart.

Two smaller ones: the same "High-protein diet" query returned values up to 9 on one download and all zeros a minute later, so I dropped it as too thin to trust. And "gym" hit 100 in August 2016, most likely Pokémon Go, which uses gyms in its game; one word, two meanings, same problem as protein.

Research log: where I looked, and what didn't work

The final story is short. Getting there took 59 Google Trends downloads covering 62 search terms, 7,641 data points, 8 analysis methods and 12 hypotheses. These are the routes that went nowhere, kept here because each one shaped the next step.

• SFA food supply data. My first try at measuring consumption. It shows how much food enters Singapore, not what people eat, and it didn't track search at all. I dropped it for the National Nutrition Survey, which measures intake directly.
• Archived supermarket pages. I hoped to count protein products on FairPrice's shelves over time with the Wayback Machine. It had no saved copies of the pages I needed.
• Weekly data. I tried shorter time ranges to get weekly Trends data for a sharper view of the April peak. The exports came back monthly every time, so the whole study is monthly.
• A paid market report. The full Euromonitor report would have given sales by format and brand. It's priced for companies, not a personal project, so I used the figures CNA published from it.
• The wrong Oatside. My first Oatside download picked the company topic, which barely registers. The search term tells the real story. Same lesson as the Protein topic: check what Google thinks you asked for.
• South Korea. I downloaded Korea for the Asian comparison, but its protein and keto lines both collapse to near zero and spring back in the same months (July to October 2025, then October 2026). Two unrelated topics don't lose their audience together, so I treated it as a data problem and dropped the Korean topic. The Korean word for protein, 단백질, turned out to be clean, and it tells the habit story better than the topic would have.
• Japan's protein topic. My first Japan download used the Protein topic, which said Japan peaked in October 2022 and fell to 37% of that within 15 months. Then the Japanese words told a different story: the powder word peaked in June 2020 and never fell that far, and neither word shows the topic's sudden jump in early 2021. I trusted the words, which are what people actually typed, and rewrote the Japan finding.
• AI chatbots. Some people now ask an AI chatbot to plan meals and track workouts instead of searching Google. That interest is invisible in Trends, and there's no open data on it yet, so I note it as a limit rather than measure it.

What I'd do next

Search data tells me what people pay attention to. The next questions need other data.

• Is protein following low fat's arc? Low fat claims dominated food launches for decades before fading. USDA FoodData Central lists branded products with launch dates, so I could chart the share of new products with protein in the name, year by year, and see whether the claims are still climbing.
• Do product launches line up with the spikes? The January to April 2026 jump might match big launches and campaigns. That needs a dated list of launches set against the monthly series.
• Will Singapore follow the US cool down? My prediction: within 6 to 12 months. Each month I'll check interest against the same month last year, the 3 month against the 12 month average, and the Bollinger z-score, and score the prediction against what happens.
• Does TikTok show up in Google searches? TikTok's research API isn't open to me, but its Creative Center shows a few months of hashtag trends. Logging the top food hashtags each month would build a dataset over time.
• Who is behind the trend? Age and gender need a survey, not search data.

Reflection

I'd been learning Python on a data course for months without anything to show for it. This project was where the course finally became useful: moving averages, Bollinger bands, seasonality and correlation all came straight from class, applied to a question I actually cared about.

The part I'll remember is how often my gut was wrong. I expected a January spike and found a December slump. I guessed protein would last until the next trend overtook it, and when I tested fibre as that next trend, the data said no successor was visible. Writing my expectations down before each chart made those misses easy to see, and harder to quietly explain away.

Methods and sources

Tools: Python (pandas, NumPy, scikit-learn, statsmodels, Lets-Plot) in Google Colab.

Data: Google Trends, Singapore, Web Search, monthly, January 2004 to October 2026, downloaded by hand on 8 November 2026. Each comparison term was downloaded alongside Protein (Nutrient topic) as an anchor. November 2026 was dropped as a partial month. Analysis runs from January 2019. Matching downloads of Protein and Ketogenic diet (both topics) for the US, Japan, Hong Kong, Malaysia and South Korea were used for the country comparison.

Other data: HPB National Nutrition Survey 2025 (protein intake, 2019, 2022 and 2025, self-reported); MOH National Population Health Survey 2025 (strength training, 2020 to 2025); sports protein sales from Euromonitor International as reported by CNA. Separate Trends downloads for 32 trend shapes, protein and oat milk brands, fitness activities and the causal impact controls. Brand downloads with more than five terms may span two downloads, so I compared their shapes only.

Methods:

• Momentum: 3 month vs 12 month rolling averages, and year on year comparison
• Hype vs normal range: 12 month Bollinger bands at 2 standard deviations, with rolling z-scores
• Seasonality: each month divided by a centred 12 month average, 2019 to 2024
• Relationships: correlations on levels, month on month changes and year on year changes, with a 0 to 12 month lag test for Ozempic
• Cross-country timing: each country scaled to its own peak since 2019
• Shape test: k-means clustering (k of 4, n_init 50) on 32 trends, each smoothed over 3 months and divided by its own peak
• Forecasting: lag features, a train/test split by date (train 2019 to 2025, test 2026), naive, seasonal naive, linear regression and random forest, scored by mean absolute error
• Causal impact: ordinary least squares on control trends, with 95% prediction intervals from statsmodels, run with and without fibre

Sources:

• The high protein craze: Fad or fixture in the Singaporean diet?, CNA, 8 May 2026
• Google Trends
• National Nutrition Survey 2022 findings, Ministry of Health
• National Nutrition Survey 2025 report, Health Promotion Board
• National Population Health Survey 2025, Ministry of Health
• Interactive scroll story for this project