Sosyal Medyada Trend Analizi Nasıl Yapılır?

What does trend analysis show?

Social media trend analysis examines how posts around a topic or brand change over time. The goal is not to rank the most-used hashtag — the real questions are: when did the rise begin, is it lasting or a one-day spike, which sub-topics came with it, and did the extra volume translate into engagement?

Designing the analysis properly

The work you do before pulling any data determines the quality of the analysis more than the data itself does.

1. Write down the research question. "What are people saying about the brand?" is not a question. "Did brand conversation volume rise after the product announcement, and if so for how long?" is.
2. Fix the scope. Date range, language, country, spelling variants of the brand name, included and excluded keywords. If your brand name is also an ordinary word (an animal or a city, say), the exclusion list is half the analysis.
3. Choose a comparison baseline. Equal-length windows before and after the announcement, the same week last year, or a similar control topic. Without a baseline, the sentence "it went up" means nothing.
4. Build a data dictionary. Which field comes from which platform, how is engagement defined, what is the time zone?
5. Verify the result by hand. Read 50 random posts. That is the only way you will notice your search query is picking up off-topic content.

Data structure and first preparation

At minimum your analysis table needs a timestamp and the text for every post. Engagement fields may be missing or defined differently depending on the platform.

import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

df = pd.read_csv("social_data.csv")

df["created_at"] = pd.to_datetime(
    df["created_at"], utc=True, errors="coerce"
)
df = df.dropna(subset=["created_at"]).copy()

# Convert to local time — critical for hour-of-day analysis
df["created_local"] = df["created_at"].dt.tz_convert("Europe/Istanbul")
df["day"] = df["created_local"].dt.floor("D")

print("Records:", len(df),
      "| Users:", df["author_id"].nunique(),
      "| Range:", df["day"].min(), "→", df["day"].max())

Skipping the time zone conversion introduces a systematic error in hour-based analysis. A post at midnight UTC is three in the morning Turkish time — you would answer "what hour do people talk most?" with a three-hour shift.

Conversation volume and change

The most basic indicator is the daily post count. The raw daily line is jumpy; a moving average smooths it. Plot both — if you show only the moving average you have hidden the spike.

daily = (df.groupby("day")
           .agg(posts=("text", "size"),
                users=("author_id", "nunique")))

daily["avg_7d"] = daily["posts"].rolling(7).mean()

sns.set_theme(style="whitegrid")
fig, ax = plt.subplots(figsize=(11, 5))
ax.plot(daily.index, daily["posts"], alpha=.4, label="Daily posts")
ax.plot(daily.index, daily["avg_7d"], linewidth=2.5, label="7-day average")
ax.set(title="Daily conversation volume", xlabel="Date", ylabel="Posts")
ax.legend(); fig.tight_layout(); plt.show()

Computing period change

daily["prev"] = daily["posts"].shift(1)

daily["change_pct"] = (
    (daily["posts"] - daily["prev"])
    / daily["prev"].replace(0, pd.NA) * 100
)

print(daily.tail(7).round(1))

If the previous period is zero, the percent change is undefined; do not display an infinite ratio. More importantly: at low volume, percentages mislead. Going from 2 posts to 6 is a 200% increase and means nothing. Always put the absolute number next to the percentage.

Popular topics and hashtags

Hashtag frequency gives a quick clue about what is being discussed. Because the same tag can appear several times in one post, counting each tag once per post gives a more faithful distribution.

import re
from collections import Counter

def find_tags(text):
    if not isinstance(text, str):
        return set()
    # Unique tags per post
    return set(re.findall(r"#([\wğüşöçıİĞÜŞÖÇ]+)", text.lower()))

counter = Counter(
    tag for text in df["text"]
    for tag in find_tags(text)
)

top = pd.Series(dict(counter)).sort_values(ascending=False).head(10)

fig, ax = plt.subplots(figsize=(9, 5))
sns.barplot(x=top.values, y=["#" + t for t in top.index], ax=ax, color="#8b5cf6")
ax.set(title="Ten most used hashtags", xlabel="Post count", ylabel="")
fig.tight_layout(); plt.show()

Be careful with lower() on Turkish text: Python's lower() is not locale-aware, so it maps a capital "I" to a dotted "i" rather than the dotless "ı". If your brand name contains a capital I, account for that and use a manual mapping where needed.

Engagement trends

Examine engagement from two angles: daily total (how much interest overall) and average per post (how well the content lands). The two do not have to move together, and where they diverge is usually the most interesting finding.

engagement_cols = ["likes", "replies", "shares"]
df[engagement_cols] = df[engagement_cols].fillna(0)
df["engagement"] = df[engagement_cols].sum(axis=1)

by_day = df.groupby("day").agg(
    posts=("text", "size"),
    total=("engagement", "sum"),
    mean=("engagement", "mean"),
    median=("engagement", "median"),
)

# Weighted rate where impression data exists
if "impressions" in df.columns:
    valid = df["impressions"].gt(0)
    rate = (df.loc[valid, "engagement"].sum()
            / df.loc[valid, "impressions"].sum())
    print(f"Engagement/impression rate: {rate:.2%}")

Note that I added the median to the table. One viral post drags the mean entirely; the median shows how a typical post performs. If the mean is 450 and the median is 3, that day's engagement came from a single post — and that is the actual headline of the report.

When computing rates, watch the scope of the denominator: if you used only rows that have impression data, do not present the rate as if it applied to the whole dataset. Platforms define engagement differently; do not treat it as the same measure across platforms.

Interpretation: what to watch for

That last row matters especially and is missed often. A day that drops to zero on your volume chart may not be "nobody talked" but "we could not collect". The two look identical on a chart. So track collection coverage as its own series: how many API calls were made each day, how many records came back?

A repeatable workflow

1. Collect: from a permitted source, over a window that fits the question; record the collection coverage too.
2. Clean: standardise timestamps, examine duplicates and suspected bot accounts.
3. Classify: build a topic dictionary and verify samples by hand.
4. Measure: volume, unique users, period change, moving average, engagement.
5. Visualise: a line for the time series, a horizontal bar for topic ranking, a heatmap for the hour-of-day pattern.
6. Interpret: compare changes against events, write down the alternative explanations.
7. Report: keep the code, the filters, the date range and the metric formulas; update through the same flow.

Conclusion

Social media trend analysis is more than turning conversation into numbers. A good analysis does four things at once: it defines volume precisely (posts or users?), separates temporal change from noise, reads topics in context, and measures engagement with a fair denominator.

When writing the report, make a habit of including five things: the metric definition, the data coverage, the chart, plausible alternative explanations, and what you do not know. The last is the most valuable — a sentence like "this data covers only Turkish-language posts and only one platform" keeps the chart from projecting more certainty than it has.