Data Visualization Principles at (un)Common Logic

Good visualization takes messy, excessive‑extent files and makes a selection obtrusive. That is the bar we set at (un)Common Logic. We construct charts and dashboards no longer to entertain, yet to inform designated actions: wherein to invest, what to restore, which scan to run subsequent, and whilst to stop doing one thing that now not will pay off. Over time we've found out that a handful of principled conduct constantly separate visuals that drive result from those who quietly acquire mud in a bookmarks folder.

Start with the determination, no longer the dataset

Most terrible charts start out with a dump of the entirety that became hassle-free to drag. That means tempts you into plotting no matter what is shut handy rather than what the resolution on the contrary requires. We commence by way of naming two things in undeniable language: the determination at stake, and the time-frame for that determination. Example: should always we amplify our paid search budget for non‑brand queries over a higher region. With that frame, the chart virtually chooses itself. You need a view of marginal payment in line with acquisition in opposition t capacity, no longer a college of impressions, click on‑due to cost, and gadget splits.

I retain a scrap computer of questions stakeholders in actual fact ask. The entries are unglamorous and concrete: why did Tuesday sink, which audience is cannibalizing organic, wherein did the margin cross after the promo. When the query is factual, visible necessities sharpen. For a Tuesday dip, you desire a time collection with day‑of‑week banding and an annotation for a site match. For cannibalization, you desire edge‑via‑facet listed sequence to teach relative stream, now not absolute totals that confuse scale. For margin erosion, a waterfall chart throughout payment components is greater sincere than a pie.

If you can't write the resolution in a unmarried sentence, you should not organized to layout the visible.

Define the metric like a contract

At (un)Common Logic we deal with metric definitions https://www.uncommonlogic.com/about/ as model‑managed agreements. You can decide upon the precise chart and still fail if numerous groups compute the metric another way. Even a fundamental magnitude like conversion fee can mean sessions‑to‑orders, customers‑to‑orders, or clicks‑to‑leads. One patron came to us with 3 dashboards exhibiting 3 special conversion fees for the same crusade. Each changed into just right on its personal terms, and jointly dead.

We positioned the definition at the canvas. Not in a tooltip, now not buried in documentation. If the metric is a composite, we present the inputs and the calculation in a short subtitle or footnote. If sampling, filtering, or attribution legislation observe, we disclose it. We also pin the denominator to the axis label whilst the menace of misinterpretation is high. A y‑axis that reads Orders in line with 1,000 classes is unique and stops a parade of Slack questions.

Precision beats secret. Rounding can disguise development or overstate innovations. As a default, we prevent one to two decimals for prices and basis factors for very small differences after they count number. For currency, we fit the viewers. Finance desires cents. An government would possibly basically care approximately entire funds. The trick is to be steady throughout a view. Mixing models is a fast path to confusion.

Context isn't optional

A single sparkline without historical past can tell any tale you choose. We most of the time anchor charts with a baseline or a benchmark. That may very well be final era, a rolling median, a aim line, or an external index. Context makes stream meaningful. A 12 % improvement expense sounds wonderful till you observe the type grew 20 percent.

Comparisons work prime whilst they're adjacent and aligned. Put series at the equal scale if that you may. If you need to use a secondary axis, colour it conscientiously and give a boost to the mapping with labels on the collection, no longer just the legend. We additionally like small multiples while the function is to examine styles across segments. Twelve thin, equal panels beat a unmarried cluttered plot with twelve colored traces that pass like spaghetti.

Annotations deserve extra use. We mark the day a price exchange went reside, the date a tracking restore deployed, the span of a vacation. These notes do extra than give an explanation for variance. They save hours of conferences. The audience sees the intent and end result and moves directly to what to do approximately it.

Choose the simplest model that solutions the question

There is no prize for novelty. Fancy visuals are the desirable choice simply while the clear-cut one should not raise the weight. Over time, a few styles have earned permanent spots in our toolkit. They are boring and noticeably wonderful.

• Time collection with regard bands for seasonality, exhibiting cutting-edge period against a baseline
• Indexed comparisons that get started extraordinary sequence at 100 to teach relative growth
• Waterfall charts to disaggregate modification from one overall to another
• Bar charts taken care of by means of fee for rank and distribution
• Scatterplots with a trend line and quadrants for portfolio decisions

The possibility with intricate visuals is that they create cognitive overhead. If you want a legend longer than two traces, stop. If you want a paragraph to explain the best way to learn it, quit. The audience should always decode structure and coloration in seconds. Reserve not easy bureaucracy for explorations, not for the very last communication.

Use color to encode that means, now not to decorate

Color is a software, not a palette to expose off flavor. We adopt a cussed default: grey for context, a single strong coloration for recognition. When we add a 2d colour, it really is to encode a 2d size of which means, no longer to decorate the web page. The most trouble-free misuse we see is a rainbow of segments in a bar chart the place rank subjects extra than hue. That forces the mind to do extra work.

Accessibility is non‑negotiable. Around eight p.c of fellows and a smaller proportion of ladies have a few variety of coloration imaginative and prescient deficiency. We scan with a simulator and steer clear of purple‑inexperienced choices. Blue‑orange is broadly speaking safer. We do not depend upon color on my own to signal outliers or states. Line variety, dot structure, and direct labels lend a hand. High comparison among textual content and history improves clarity for all and sundry.

Legibility is a part of colour practice. Saturated fills can disguise gridlines and wash out labels. Pastels look revolutionary, however they may fail in a projector or in a screenshot compressed for e-mail. We try our palettes in grayscale to determine if the message nevertheless works. If the tale falls apart with out color, the encoding become fragile to start with.

Label with an editor’s precision

Labels turn documents into statements. Direct labeling, wherein the price or identify sits subsequent to the road or bar, outperforms a legend in such a lot situations. You do away with eye travel and reduce error. Legends belong in exploratory methods where the user demands flexibility. For a narrative chart, marketing consultant the reader.

We cut nonessential ink. Axis ticks are sparse and significant. Data labels occur solely for peaks, troughs, and the so much significant points. We circular with cause. For busy plots, we tutor totals and allow the size do the leisure. Titles do not mumble. They inform the point: Mobile CPA fell underneath target after bid caps. A useful name frees the viewer from attempting to find a moral.

Footnotes count. If there is a rationale a importance is lacking, a matter is cut back, or a spike is a generic artifact, we say so. That variety of honesty prevents chasing ghosts.

Respect scale, share, and zero

Nothing will erode trust quicker than a compressed y‑axis that turns noise into narrative. When the variable is a quantity the place zero has that means, consist of zero. Revenue, orders, spend, and headcount stay in that classification. For charges and indices, zero would possibly not be the anchor. A soar cost swap of 3 features appears to be like flat on a 0 to a hundred scale across a small panel. In that case, exhibit replace as a separate bar or a secondary small sparkline.

Log scales have their location, primarily for knowledge that spans orders of magnitude, like key-word quantity or web page load times with heavy tails. We label log charts in actual fact and never mix them with linear inside the comparable collection of panels. For percentage differences, stay away from the temptation to stack bars that suggest additive relationships. Percentages are ratios. Stacked p.c. bars can hide amazing shifts inside the center elements.

Proportion additionally applies to what number of visuals you cram into a unmarried view. A dashboard with 9 panels of equal weight is a hierarchy failure. If one chart is mission very important, give it 60 p.c. of the authentic estate and demote the relaxation. Visual weight needs to reflect industrial weight.

Show uncertainty

Executives like crisp solutions. Data not often offers them. We display uncertainty to construct enhanced choices. Confidence bands around forecasts, shading for incomplete days, and mistakes bars for A/B scan consequences avert optimism in take a look at. We label types with classes windows and remaining update dates. If a panel presentations estimates, we are saying estimate within the subtitle and colour labels a little bit in another way from discovered values.

Forecasts that behave well in backtests can still surprise in deployment. We include functional mannequin diagnostics off to the aspect in analyst views, like imply absolute proportion blunders over the previous few weeks. That context powers more suitable interpretation. It additionally encourages natural skepticism, which is low-priced insurance plan towards overfitting a story to a single chart.

Build for the proper altitude

A single source of actuality does no longer suggest a unmarried view for all of us. At (un)Common Logic we layout for three altitudes: government, supervisor, and practitioner.

Executive views compress to the fewest metrics that are expecting influence and hazard. They healthy on one screen without scrolling. Each panel is self‑explanatory and contains next steps. A spike in CAC above target triggers a callout that links to the supervisor view.

Manager perspectives focal point on allocation. They evaluate channels, merchandise, audiences, and geographies. They raise filters, yet no longer too many. We decide on slices that impression budgets, staffing, or roadmaps. A appropriate manager view helps resolution what to do this week.

Practitioner perspectives are instruments, not reviews. They reply how and why. Controls get heavy right here considering that the user necessities to isolate cohorts, examine hypotheses, and debug anomalies. We construct these with the idea that the viewer knows the tips version. That offers us room for technical labels, reference tables, and uncooked counts.

The mistake is at hand an govt a practitioner instrument, or handy a practitioner a self-esteem precis. Fit beats uniformity.

Reduce friction within the workflow

A amazing chart that takes forty seconds to load will die. We plan for latency. Pre‑aggregation, caching, and limiting default date stages avoid dashboards snappy. For high‑cardinality dimensions like queries or merchandise, we index and shop rank tables with the aid of duration so we can render excellent movers quickly. When we do desire heavy queries, we load the so much wonderful panels first and fade inside the rest. Progress warning signs lower abandonment, which subjects extra than you think that.

Naming and supplier limit friction too. We use transparent folder hierarchies, strong URLs, and consistent parameter names. If a report moves, we set redirects. We also variation dashboards and identify them by intent. It is better to have Spend Efficiency Q3 than Master Dashboard v12. That background hygiene frees teams to cognizance on interpretation.

Treat motives like product features

We narrate the 1st time someone opens a dashboard. A brief explainer video, a swift guided tour, or a number of tooltip information lower the discovering curve. Not all and sundry reads documentation. We layout the 1st‑run event like a product. Then we assess analytics to work out the place customers drop off. If most clients not ever scroll to the bottom panel, we reconsider the order or minimize it.

We also use onboarding to set norms. For illustration, we state that incomplete days are shaded and excluded from week‑over‑week comparisons until noon nearby time. That one sentence prevents a ordinary around of fake alarms each morning.

Know whilst not to visualize

Some details learn more effective as a sentence than as a chart. A benchmark like Industry CPC for non‑emblem rose 9 to 12 p.c over the last six months across top networks is a unmarried line that beats a dense column chart for lots audiences. The medium needs to serve the message. We in many instances write one or two traces straight in a dashboard above a panel to summarize the takeaways. Good annotation spares the staff from examining the image chilly.

There are other circumstances to pass the visual. If the sample length is just too small to assist a pattern, say so. Do now not plot a line for 3 data features. If the source details is in flux and seemingly to amendment materially, maintain again until eventually the system stabilizes. A unsuitable chart, obvious at the wrong time, can linger in reminiscence longer than a correction.

A transient vignette: the multi‑contact muddle

A retail customer got here to (un)Common Logic with a basic worry. Email, paid social, and organic and natural have been all claiming credit for an uptick in sales after a spring marketing campaign. Each staff had a chart that proved its case. Each chart used a diversified attribution type. Meetings grew anxious and circular. We all started with a precept that earned buy‑in from every person in the room: each one sort solutions a different query, so we are going to reveal them aspect by using side and label the query, now not the form.

We constructed 3 panels. The first confirmed remaining‑click profit by way of channel with a transparent identify, Who closed the sale. The second showed position‑situated gross sales, Who introduced and supported. The third showed an incrementality estimate from geo‑carry tests, Who moved gross sales that would not have passed off otherwise. We aligned scales, used the comparable gray context and one focal point colour consistent with panel, and annotated the era with promo dates and site outages.

Two issues happened. The groups stopped arguing about whose chart used to be the excellent one considering the questions had been clear. And the executive sponsor may possibly now make a decision grounded in exchange‑offs. Paid social did not shut many gross sales, but it played a effective assist position and showed triumphant lift in take a look at markets. We increased its funds with guardrails. Email stored credit score for closing and centred on send timing to keep away from cannibalizing organic. One visualization set, constructed on sincere concepts, created alignment without drama.

Quantify modification responsibly

Percentages play tips. A bounce from 1 to two percent is a one hundred p.c raise and nevertheless will possibly not depend to the industry. We anchor p.c ameliorations to absolute impression. A callout that announces Signup price up 0.8 points, four hundred greater signups final week, actions the room sooner than Up sixty seven p.c. For fiscal metrics, we exhibit adjustments in cash in which simple. Framing things for the reason that individuals make portfolio choices with limited budgets, not with limitless urge for food for p.c. beneficial properties.

We additionally discourage stacked comparisons across mismatched totals. Comparing click on‑with the aid of rates for two advertising with assorted impressions is superb, however stacking these bars can indicate the identical base. We prefer facet‑with the aid of‑side bars with direct labels and explicit base counts in a footnote. If a metric will probably be gamed by using replacing the denominator, we name it out and steadily pair it with a balancing metric. For illustration, we show settlement per upload‑to‑cart along payment consistent with purchase to expose funnel friction.

Keep exploration break free presentation

Analysts desire room to play. Executives want crisp views. Mixing the 2 creates artifacts like 40 filters, 8 legends, and a monitor that reads like a cockpit. We separate exploration from presentation. The exploration lives in notebooks and sandbox dashboards the place we take a look at hypotheses and iterate in a timely fashion. When a tale is set, we promote a sparkling variant to the presentation layer with the fewest controls had to drill into movement. That separation additionally speeds load time and eases protection.

We deal with the start from exploration to presentation as a release. We freeze metric definitions, rfile inputs, and run person exams with a handful of true stakeholders. Feedback loops are swift. We may slightly ship a minimum, solid view and extend than overthrow a crowded format that nobody enormously trusts.

A short chart preference map for primary questions

• How is efficiency trending: a time sequence with a baseline band, plus a small varied for key segments
• Where did the amendment come from: a waterfall among two totals, with elements sorted by using contribution
• What deserve to we prioritize: a scatterplot with affect on the x‑axis, attempt or money on the y‑axis, and bubble size for volume
• Which editions are successful: a bar chart with trust intervals, ordered with the aid of uplift
• Are we cannibalizing: indexed traces start at a hundred for overlapping merchandise or channels

Small information that raise weight

We sweat information that sound fussy until they store 1 / 4. Here are just a few that recur.

Time zones: pick out one in step with dashboard and print it near the name. Mixed zones quietly ruin comparisons.

Partial periods: coloration them and exclude them with the aid of default from comparisons. If you embrace them, say why.

Week definitions: some groups run Sunday to Saturday, others Monday to Sunday. Set a rule and stick with it.

Currency: train the foreign money symbol, and in case you mixture currencies throughout areas, convert or separate views. An unlabeled buck is an mistakes waiting to manifest.

Index starts offevolved: define your index anchor definitely. If you say Day zero equals marketing campaign release, ensure each collection starts there.

Tiny law ward off massive error.

Performance and scale with no drama

Charts need to carry out less than load. We check with real production volumes, not toy samples. If a plot fails with 20 million rows, it fails, era. We build rollups at every single day or weekly grains for ancient views and hold raw, high‑granularity archives behind drill‑throughs in which in simple terms analysts move. We prune pricey transforms out of the live layer. When a calculation is steady and used greatly, we materialize it.

We additionally plan for snapshots. Historical accuracy concerns in advertising and marketing and product analytics. If a accomplice retroactively fixes attribution or a feed reprocesses, you can come to be with transferring goals. We snapshot day-after-day aggregates so the previous remains put. Reproducibility is a consumer journey feature, even if the person in no way sees the machinery.

Testing visuals like features

A visualization is a product in miniature, so we take a look at it. We run hallway assessments with 3 to five individuals who were not worried in construction it. Each has 60 seconds to tell us what the chart is saying and what they could do next. If their solution diverges from rationale, we modify labels, scales, or type. That inexpensive scan catches troubles in the past they calcify.

We also video display usage. A dashboard that no person opens is just not a luck. We log perspectives, live time, scroll intensity, and universal clear out combinations. If a panel by no means gets consideration, we ask why. Maybe it belongs in the practitioner view. Maybe it should depart. Ruthless pruning helps to keep the sign powerful.

A realistic assessment checklist before you ship

• State the determination and the timeframe in the name or subtitle
• Put metric definitions and denominators at the canvas, not just in docs
• Check shade comparison, and be certain the story holds in grayscale
• Verify axis picks, label right away, and keep legends while possible
• Annotate favourite activities, demonstrate uncertainty, and shade incomplete periods

Governance with no bureaucracy

At (un)Common Logic, governance manner shared requisites that make collaboration more easy, now not red tape that slows work. We preserve a lightweight trend e-book with examples, shade palettes, typography suggestions, and hottest chart paperwork. We keep it with living code snippets for straightforward visuals so analysts can compile constant charts briskly. New staff participants analyze turbo, and stakeholders do now not have got to relearn the language of knowledge on every occasion the author differences.

We also audit dashboards quarterly. The audit isn't always approximately blame. It is ready health. We ask whether the dashboard nonetheless answers the judgements it was once equipped for, even if metrics have drifted, and whether controls match the modern org. Sunsetting is a match exercise. Every retirement is a small reward of consciousness to come back to the teams.

Ethics and honesty

Visuals lift chronic. They can nudge selections, build power, and create self belief. With that comes duty. We do no longer cover caveats in footnotes when stakes are prime. We do no longer compress scales to dramatize flat tendencies. We do no longer cherry‑elect durations to flatter a campaign. We would pretty bring dangerous news cleanly than lengthen a corrective motion. That ethic earns have faith, and have confidence maintains stakeholders coming back to the details even if it hurts.

One behavior enables: show the alternative view. If there is a attainable 2nd interpretation, encompass it along the commonplace. That disarms allegations of bias and types the reasonably guilty skepticism we want throughout the business enterprise.

Closing thought

The easiest compliment a visualization can earn is short and direct: this helped me pick. Getting there may be much less approximately creative aptitude and extra about a sequence of disciplined decisions. Start with the choice. Define the metric. Add context. Choose the only shape. Use colour with motive. Label and annotate with care. Respect uncertainty. Fit the view to the altitude. Reduce friction. Test like a product. Govern frivolously and ethically.

At (un)Common Logic we go back to those principles when you consider that they work. They velocity selections, diminish noise, and turn details into a partner in place of a puzzle. And when a stakeholder opens a dashboard on a busy morning, sees a clean story, and knows what to do subsequent, the entire quiet work behind the curtain turned into price it.